Memory interleaving in a high-speed switching environment

Abstract

In one embodiment of the present invention, a system for memory interleaving in a high-speed switching environment includes multiple memory units that each include one or more memory devices. The system also includes multiple port modules. Each port module can receive a packet communicated from a component of a communications network, write the received packet to one or more of the memory units, and read a packet from one or more of the memory units for communication to the component of the communications network. The system also includes an interconnection network including a hierarchical structure that includes one or more switching stages. The interconnection network couples the memory units to the port modules such that each of the port modules can write to each of the memory units according to a first schedule and read from each of the memory units according to a second schedule and such that a first port module can read a first portion of a packet from one or more memory units for communication to a first component of the communications network before a second port module has received a second portion of the packet communicated from a second component of the communications network.

Description

TECHNICAL FIELD OF THE INVENTION

[0001] This invention relates generally to communication systems and more particularly to memory interleaving in a high-speed switching environment.

BACKGROUND OF THE INVENTION

[0002] High-speed serial interconnects have become more common in communications environments, and, as a result, the role that switches play in these environments has become more important. Traditional switches do not provide the scalability and switching speed typically needed to support these interconnects.

SUMMARY OF THE INVENTION

[0003] Particular embodiments of the present invention may reduce or eliminate disadvantages and problems traditionally associated with memory in a high-speed switching environment.

[0004] In one embodiment of the present invention, a system for memory interleaving in a high-speed switching environment includes multiple memory units that each include one or more memory devices. The system also includes multiple port modules. Each port module can receive a packet communicated from a component of a communications network, write the received packet to one or more of the memory units, and read a packet from one or more of the memory units for communication to the component of the communications network. The system also includes an interconnection network including a hierarchical structure that includes one or more switching stages. The interconnection network couples the memory units to the port modules such that each of the port modules can write to each of the memory units according to a first schedule and read from each of the memory units according to a second schedule and such that a first port module can read a first portion of a packet from one or more memory units for communication to a first component of the communications network before a second port module has received a second portion of the packet communicated from a second component of the communications network.

[0005] Particular embodiments of the present invention provide one or more advantages. Particular embodiments reduce memory requirements associated with multicast traffic. In particular embodiments, port modules share memory resources, which tends to eliminate head-of-line blocking, reduce memory requirements, and enable more efficient handling of changes in load conditions at port modules. Particular embodiments provide cut-through forwarding, which provides one or more advantages over store-and-forward techniques. Particular embodiments provide delayed cut-through forwarding, which also provides one or more advantages over store-and-forward techniques. Particular embodiments increase the throughput of a switch core. Particular embodiments increase the speed at which packets are switched by a switch core. Particular embodiments reduce the fall-through latency of a switch core, which is important for cluster applications. Particular embodiments are embodied in a single integrated circuit (IC), or chip. Particular embodiments reduce the power dissipation of a switch core. Particular embodiments can be used in different applications, such as Ethernet switches, INFINIBAND switches, 3GIO switches, HYPERTRANSPORT switches, RAPID IO switches, or proprietary backplane switches. Certain embodiments provide all, some, or none of these technical advantages, and certain embodiments provide one or more other technical advantages readily apparent to those skilled in the art from the figures, descriptions, and claims included herein.

BRIEF DESCRIPTION OF THE DRAWINGS

[0006] To provide a more complete understanding of the present invention and the features and advantages thereof, reference is made to the following description, taken in conjunction with the accompanying drawings, in which:

[0007] FIG. 1 illustrates an example system area network;

[0008] FIG. 2 illustrates an example switch of a system area network;

[0009] FIG. 3 illustrates an example switch core of a switch;

[0010] FIG. 4 illustrates an example stream memory of a switch core logically divided into blocks;

[0011] FIG. 5 illustrates example scheduling at two switching units of a switch core for write operations to three memory banks by six port modules;

[0012] FIG. 6 illustrates example scheduling at a switching unit of a switch core for read operations from twenty-four memory units by a port module;

[0013] FIG. 7 illustrates an example memory bank of a switch core;

[0014] FIG. 8 illustrates example scheduling at three bank switching units of a memory bank for read operations to two memory units via four switching units;

[0015] FIG. 9 illustrates example scheduling for write operations to and read operations from eight memory units of a memory bank via four switching units; and

[0016] FIG. 10 illustrates an example method for memory interleaving in a high-speed switching environment.

DESCRIPTION OF EXAMPLE EMBODIMENTS

[0017] FIG. 1 illustrates an example system area network 10 that includes a serial or other interconnect 12 supporting communication among one or more server systems 14; one or more storage systems 16; one or more network systems 18; and one or more routing systems 20 coupling interconnect 12 to one or more other networks, which include one or more local area networks (LANs), wide area networks (WANs), or other networks. Server systems 14 each include one or more central processing units (CPUs) and one or more memory units. Storage systems 16 each include one or more channel adaptors (CAs), one or more disk adaptors (DAs), and one or more CPU modules (CMs). Interconnect 12 includes one or more switches 22, which, in particular embodiments, include Ethernet switches, as described more fully below. The components of system area network 10 are coupled to each other using one or more links, each of which includes one or more computer buses, local area networks (LANs), metropolitan area networks (MANs), wide area networks (WANs), portions of the Internet, or other wireline, optical, wireless, or other links. Although system area network 10 is described and illustrated as including particular components coupled to each other in a particular configuration, the present invention contemplates any suitable system area network including any suitable components coupled to each other in any suitable configuration.

[0018] FIG. 2 illustrates an example switch 22 of system area network 10. Switch 22 includes multiple ports 24 and a switch core 26. Ports 24 are each coupled to switch core 26 and a component of system area network 10 (such as a server system 14, a storage system 16, a network system 18, a routing system 20, or another switch 22). A first port 24 receives a packet from a first component of system area network 10 and communicates the packet to switch core 26 for switching to a second port 24, which communicates the packet to a second component of system area network 10. Reference to a packet can include a packet, datagram, frame, or other unit of data, where appropriate. Switch core 26 receives a packet from a first port 24 and switches the packet to one or more second ports 24, as described more fully below. In particular embodiments, switch 22 includes an Ethernet switch. In particular embodiments, switch 22 can switch packets at or near wire speed.

[0019] FIG. 3 illustrates an example switch core 26 of switch 22. Switch core 26 includes twelve port modules 28, stream memory 30, tag memory 32, central agent 34, and routing module 36. The components of switch core 26 are coupled to each other using buses or other links. In particular embodiments, switch core 26 is embodied in a single IC. In a default mode of switch core 26, a packet received by switch core 26 from a first component of system area network 10 can be communicated from switch core 26 to one or more second components of system area network 10 before switch core 26 receives the entire packet. In particular embodiments, cut-through forwarding provides one or more advantages (such as reduced latency, reduced memory requirements, and increased throughput) over store-and-forward techniques. Switch core 26 can be configured for different applications. As an example and not by way of limitation, switch core 26 can be configured for an Ethernet switch 22 (which includes a ten-gigabit Ethernet switch 22 or an Ethernet switch 22 in particular embodiments); an INFINIBAND switch 22; a 3GIO switch 22; a HYPERTRANSPORT switch 22; a RAPID IO switch 22; a proprietary backplane switch 22 for storage systems 16, network systems 18, or both; or other switch 22.

[0020] A port module 28 provides an interface between switch core 26 and a port 24 of switch 22. Port module 28 is coupled to port 24, stream memory 30, tag memory 32, central agent 34, and routing table 36. In particular embodiments, port module 28 includes both input logic (which is used for receiving a packet from a component of system area network 10 and writing the packet to stream memory 30) and output logic (which is used for reading a packet from stream memory 30 and communicating the packet to a component of system area network 10). As an alternative, in particular embodiments, port module 28 includes only input logic or only output logic. Reference to a port module 28 can include a port module 28 that includes input logic, output logic, or both, where appropriate. Port module 28 can also include an input buffer for inbound flow control. In an Ethernet switch 22, a pause function can be used for inbound flow control, which can take time to be effective. The input buffer of port module 28 can be used for temporary storage of a packet that is sent before the pause function stops incoming packets. Because the input buffer would be unnecessary if credits are exported for inbound flow control, as would be the case in an INFINIBAND switch 22, the input buffer is optional. In particular embodiments, the link coupling port module 28 to stream memory 30 includes two links: one for write operations (which include operations of switch core 26 in which data is written from a port module 28 to stream memory 30) and one for read operations (which include operations of switch core 26 in which data is read from stream memory 30 to a port module 28). Each of these links can carry thirty-six bits, making the data path between port module 28 and stream memory 30 thirty-six bits wide in both directions.

[0021] A packet received by a first port module 28 from a first component of system area network 10 is written to stream memory 30 from first port module 28 and later read from stream memory 30 to one or more second port modules 28 for communication from second port modules 28 to one or more second components of system area network 10. Reference to a packet being received by or communicated from a port module 28 can include the entire packet being received by or communicated from port module 28 or only a portion of the packet being received by or communicated from port module 28, where appropriate. Similarly, reference to a packet being written to or read from stream memory 30 can include the entire packet being written to or read from stream memory 30 or only a portion of the packet being written to or read from stream memory 30, where appropriate. Any port module 28 that includes input logic can write to stream memory 30, and any port module 28 that includes output logic can read from stream memory 30. In particular embodiments, the sharing of stream memory 30 by port modules 28 eliminates head-of-line blocking (thereby increasing the throughput of switch core 26), reduces memory requirements associated with switch core 26, and enables switch core 26 to more efficiently handle changes in load conditions at port modules 28.

[0022] Stream memory 30 of switch core 26 is logically divided into blocks 38, which are further divided into words 40, as illustrated in FIG. 4. A row represents a block 38, and the intersection of the row with a column represents a word 40 of block 38. In particular embodiments, stream memory 30 is divided into 1536 blocks 38, each block 38 includes twenty-four words 40, and a word 40 includes seventy-two bits. Although stream memory 30 is described and illustrated as being divided into a particular number of blocks 38 that are divided into a particular number of words 40 including a particular number of bits, the present invention contemplates stream memory 30 being divided into any suitable number of blocks 38 that are divided into any suitable number of words 40 including any suitable number of bits. Packet size can vary from packet to packet. A packet that includes as many bits as or fewer bits than a block 38 can be written to one block 38, and a packet that includes more bits than a block 38 can be written to more than one block 38, which need not be contiguous with each other.

[0023] When writing to or reading from a block 38, a port module 28 can start at any word 40 of block 38 and write to or read from words 40 of block 38 sequentially. Port module 28 can also wrap around to a first word 40 of block 38 as it writes to or reads from block 38. A block 38 has an address that can be used to identify block 38 in a write operation or a read operation, and an offset can be used to identify a word 40 of block 38 in a write operation or a read operation. As an example, consider a packet that is 4176 bits long. The packet has been written to fifty-eight words 40, starting at word 40f of block 38a and continuing to word 40k of block 38d, excluding block 38b. In the write operation, word 40f of block 38a is identified by a first address and a first offset, word 40f of block 38c is identified by a second address and a second offset, and word 40f of block 38d is identified by a third address and a third offset. The packet can also be read from stream memory 30 starting at word 40f of block 38a and continuing to word 40k of block 38d, excluding block 38b. In the read operation, word 40f of block 38a can be identified by the first address and the first offset, word 40f of block 38c can be identified by the second address and the second offset, and word 40f of block 38d can be identified by the third address and the third offset.

[0024] Tag memory 32 includes multiple linked lists that can each be used by a first port module 28 to determine a next block 38 to which to write and by one or more second port modules 28 to determine a next block 38 from which to read. Tag memory 32 also includes a linked list that can be used by central agent 34 to determine a next block 38 that can be made available to a port module 28 for a write operation from port module 28 to stream memory 30, as described more fully below. Tag memory 32 includes multiple entries, at least some of which each correspond to a block 38 of stream memory 30. Each block 38 of stream memory 30 has a corresponding entry in tag memory 32. An entry in tag memory 32 can include a pointer to another entry in tag memory 32, resulting in a linked list.

[0025] Entries in tag memory 32 corresponding to blocks 38 that are available to a port module 28 for write operations from port module 28 to stream memory 30 can be linked together such that port module 28 can determine a next block 38 to which to write using the linked entries. As an example, consider four blocks 38 that are available to port module 28 for write operations from port module 28 to stream memory 30. A first entry in tag memory 32 corresponding to a first block 38 includes a pointer to a second block 38, a second entry in tag memory 32 corresponding to second block 38 includes a pointer to a third block 38, and a third entry in tag memory 32 corresponding to third block 38 includes a pointer to a fourth block 38. Port module 28 writes to first block 38 and, while port module 28 is writing to first block 38, uses the pointer in the first entry to determine a next block 38 to which to write. The pointer refers port module 28 to second block 38, and, when port module 28 has finished writing to first block 38, port module 28 writes to second block 38. While port module 28 is writing to second block 38, port module 28 uses the pointer in the second entry to determine a next block 38 to which to write. The pointer refers port module 28 to third block 38, and, when port module 28 has finished writing to second block 38, port module 28 writes to third block 38. While port module 28 is writing to third block 38, port module 28 uses the pointer in the third entry to determine a next block 38 to which to write. The pointer refers port module 28 to fourth block 38, and, when port module 28 has finished writing to third block 38, port module 28 writes to fourth block 38. A linked list in tag memory 32 cannot be used by more than one port module 28 to determine a next block 38 to which to write.

[0026] When a block 38 is made available to a port module 28 for write operations from port module 28 to stream memory 30, an entry in tag memory 32 corresponding to block 38 can be added to the linked list that port module 28 is using to determine a next block 38 to which to write. As an example, consider the linked list described above. If the fourth entry is the last element of the linked list, when a fifth block 38 is made available to port module 28, the fourth entry can be modified to include a pointer to fifth block 38.

[0027] A linked list in tag memory 32 that a first port module 28 is using to determine a next block 38 to which to write can also be used by one or more second port modules 28 to determine a next block 38 from which to read. As an example, consider the linked list described above. A first portion of a packet has been written from first port module 28 to first block 38, a second portion of the packet has been written from first port module 28 to second block 38, and a third and final portion of the packet has been written from first port module 28 to third block 38. An end mark has also been written to third block 38 to indicate that a final portion of the packet has been written to third block 38. A second port module 28 reads from first block 38 and, while second port module 28 is reading from first block 38, uses the pointer in the first entry to determine a next block 38 from which to read. The pointer refers second port module 28 to second block 38, and, when second port module 28 has finished reading from first block 38, second port module 28 reads from second block 38. While second port module 28 is reading from second block 38, second port module 28 uses the pointer in the second entry to determine a next block 38 from which to read. The pointer refers second port module 28 to third block 38, and, when second port module 28 has finished reading from second block 38, second port module 28 reads from third block 38. Second port module 28 reads from third block 38 and, using the end mark in third block 38, determines that a final portion of the packet has been written to third block 38. While a linked list in tag memory 32 cannot be used by more than one first port module 28 to determine a next block 38 to which to write, the linked list can be used by one or more second port modules 28 to determine a next block 38 from which to read.

[0028] Different packets can have different destinations, and the order in which packets make their way through stream memory 30 need not be first in, first out (FIFO). As an example, consider a first packet received and written to one or more first blocks 38 before a second packet is received and written to one or more second blocks 38. The second packet could be read from stream memory 30 before the first packet, and second blocks 38 could become available for other write operations before first blocks 38. In particular embodiments, a block 38 of stream memory 30 to which a packet has been written can be made available to a port module 28 for a write operation from port module 28 to block 38 immediately after the packet has been read from block 38 by all port modules 28 that are designated port modules 28 of the packet. A designated port module 28 of a packet includes a port module 28 coupled to a component of system area network 10, downstream from switch core 26, that is a final or intermediate destination of the packet.

[0029] In particular embodiments, credits are allocated to input logic of port modules 28 and are used to manage write operations. Using credits to manage write operations can facilitate cut-through forwarding by switch core 26, which reduces latency, increases throughput, and reduces memory requirements associated with switch core 26. Also, if credits are used to manage write operations, determinations regarding which port module 28 can write to which block 38 at which time can be made locally at port modules 28, which increases the throughput and switching speed of switch core 26. Using credits to manage write operations can also eliminate head-of-line blocking and provide greater flexibility in the distribution of memory resources among port modules 28 in response to changing load conditions at port modules 28. A credit corresponds to a block 38 of stream memory 30 and can be used by a port module 28 to write to block 38. A credit can be allocated to a port module 28 from a pool of credits, which is managed by central agent 34. Reference to a credit being allocated to a port module 28 includes a block 38 corresponding to the credit being made available to port module 28 for a write operation from port module 28 to block 38, and vice versa.

[0030] A credit in the pool of credits can be allocated to any port module 28 and need not be allocated to any particular port module 28. A port module 28 can use only a credit that is available to port module 28 and cannot use a credit that is available to another port module 28 or that is in the pool of credits. A credit is available to port module 28 if the credit has been allocated to port module 28 and port module 28 has not yet used the credit. A credit that has been allocated to port module 28 is available to port module 28 until port module 28 uses the credit. A credit cannot be allocated to more than one port module 28 at a time, and a credit cannot be available to more than one port module 28 at the same time. In particular embodiments, when a first port module 28 uses a credit to write a packet to a block 38 corresponding to the credit, the credit is returned to the pool of credits immediately after all designated port modules 28 of the packet have read the packet from block 38.

[0031] Central agent 34 can allocate credits to port modules 28 from the pool of credits. As an example, central agent 34 can make an initial allocation of a predetermined number of credits to a port module 28. In particular embodiments, central agent 34 can make an initial allocation of credits to port module 28 at the startup of switch core 26 or in response to switch core 26 being reset. As another example, central agent 34 can allocate a credit to a port module 28 to replace another credit that port module 28 has used. In particular embodiments, when port module 28 uses a first credit, port module 28 notifies central agent 34 that port module 28 has used the first credit, and, in response to port module 28 notifying central agent 34 that port module 28 has used the first credit, central agent 34 allocates a second credit to port module 28 to replace the first credit, but only if the number of blocks 38 that are being used by port module 28 does not meet or exceed an applicable limit. Reference to a block 38 that is being used by a port module 28 includes a block 38 to which a packet has been written from port module 28 and from which all designated port modules 28 of the packet have not read the packet. By replacing, up to an applicable limit, credits used by port module 28, the number of credits available to port module 28 can be kept relatively constant and, if the load conditions at port module 28 increase, more blocks 38 can be supplied to port module 28 in response to the increase in load conditions at port module 28. A limit can be applied to the number of blocks used by port module 28, which can prevent port module 28 from using too many blocks 38 and thereby use up too many shared memory resources. The limit can be controlled dynamically based on the number of credits in the pool of credits. If the number of credits in the pool of credits decreases, the limit can also decrease. The calculation of the limit and the process according to which credits are allocated to port module 28 can take place out of the critical path of packets through switch core 26, which increases the switching speed of switch core 26.

[0032] A linked list in tag memory 32 can be used by central agent 34 to determine a next credit that can be allocated to a port module 28. The elements of the linked list can include entries in tag memory 32 corresponding to blocks 38 that in turn correspond to credits in the pool of credits. As an example, consider four credits in the pool of credits. A first credit corresponds to a first block 38, a second credit corresponds to a second block 38, a third credit corresponds to a third block 38, and a fourth credit corresponds to a fourth block 38. A first entry in tag memory 32 corresponding to first block 38 includes a pointer to second block 38, a second entry in tag memory 32 corresponding to second block 38 includes a pointer to third block 38, and a third entry in tag memory 32 corresponding to third block 38 includes a pointer to fourth block 38. Central agent 34 allocates the first credit to a port module 28 and, while central agent 34 is allocating the first credit to a port module 28, uses the pointer in the first entry to determine a next credit to allocate to a port module 28. The pointer refers central agent 34 to second block 38, and, when central agent 34 has finished allocating the first credit to a port module 28, central agent 34 allocates the second credit to a port module 28. While central agent 34 is allocating the second credit to a port module 28, central agent 34 uses the pointer in the second entry to determine a next credit to allocate to a port module 28. The pointer refers central agent 34 to third block 38, and, when central agent 34 has finished allocating the second credit to a port module 28, central agent allocates the third credit to a port module 28. While central agent 34 is allocating the third credit to a port module 28, central agent 34 uses the pointer in the third entry to determine a next credit to allocate to a port module 28. The pointer refers central agent 34 to fourth block 38, and, when central agent 34 has finished allocating the third credit to a port module 28, central agent allocates the fourth credit to a port module 28.

[0033] When a credit corresponding to a block 38 is returned to the pool of credits, an entry in tag memory 32 corresponding to block 38 can be added to the end of the linked list that central agent 34 is using to determine a next credit to allocate to a port module 28. As an example, consider the linked list described above. If the fourth entry is the last element of the linked list, when a fifth credit corresponding to a fifth block 38 is added to the pool of credits, the fourth entry can be modified to include a pointer to a fifth entry in tag memory 32 corresponding to fifth block 38. Because entries in tag memory 32 each correspond to a block 38 of stream memory 30, a pointer that points to a block 38 also points to an entry in tag memory 32.

[0034] When a port module 28 receives an incoming packet, port module 28 determines whether enough credits are available to port module 28 to write the packet to stream memory 30. In particular embodiments, if enough credits are available to port module 28 to write the packet to stream memory 30, port module 28 can write the packet to stream memory 30 using one or more credits. In particular embodiments, if enough credits are not available to port module 28 to write the packet to stream memory 30, port module 28 can write the packet to an input buffer and later, when enough credits are available to port module 28 to write the packet to stream memory 30, write the packet to stream memory 30 using one or more credits. As an alternative to port module 28 writing the packet to an input buffer, port module 28 can drop the packet. In particular embodiments, if enough credits are available to port module 28 to write only a portion of the packet to stream memory 30, port module 28 can write to stream memory 30 the portion of the packet that can be written to stream memory 30 using one or more credits and write one or more other portions of the packet to an input buffer. Later, when enough credits are available to port module 28 to write one or more of the other portions of the packet to stream memory 30, port module 28 can write one or more of the other portions of the packet to stream memory 30 using one or more credits. In particular embodiments, delayed cut-through forwarding, like cut-through forwarding, provides one or more advantages (such as reduced latency, reduced memory requirements, and increased throughput) over store-and-forward techniques. Reference to a port module 28 determining whether enough credits are available to port module 28 to write a packet to stream memory 30 includes port module 28 determining whether enough credits are available to port module 28 to write the entire packet to stream memory 30, write only a received portion of the packet to stream memory 30, or write at least one portion of the packet to stream memory 30, where appropriate.

[0035] In particular embodiments, the length of an incoming packet cannot be known until the entire packet has been received. In these embodiments, a maximum packet size (according to an applicable set of standards) can be used to determine whether enough credits are available to a port module 28 to write an incoming packet that has been received by port module 28 to stream memory 30. According to a set of standards published by the Institute of Electrical and Electronics Engineers (IEEE), the maximum size of an Ethernet frame is 1500 bytes. According to a de facto set of standards, the maximum size of an Ethernet frame is nine thousand bytes. As an example and not by way of limitation, consider a port module 28 that has received only a portion of an incoming packet. Port module 28 uses a maximum packet size (according to an applicable set of standards) to determine whether enough credits are available to port module 28 to write the entire packet to stream memory 30. Port module 28 can make this determination by comparing the maximum packet size with the number of credits available to port module 28. If enough credits are available to port module 28 to write the entire packet to stream memory 30, port module 28 can write the received portion of the packet to stream memory 30 using one or more credits and write one or more other portions of the packet to stream memory 30 using one or more credits when port module 28 receives the one or more other portions of the packet.

[0036] A port module 28 can monitor the number of credits available to port module 28 using a counter. When central agent 34 allocates a credit to port module 28, port module 28 increments the counter by an amount, and, when port module 28 uses a credit, port module 28 decrements the counter by an amount. The current value of the counter reflects the current number of credits available to port module 28, and port module 28 can use the counter to determine whether enough credits are available to port module 28 to write a packet from port module 28 to stream memory 30. Central agent 34 can also monitor the number of credits available to port module 28 using a counter. When central agent 34 allocates a credit to port module 28, central agent 34 increments the counter by an amount, and, when port module 28 notifies central agent 34 that port module 28 has used a credit, central agent 34 decrements the counter by an amount. The current value of the counter reflects the current number of credits available to port module 28, and central agent 34 can use the counter to determine whether to allocate one or more credits to port module 28. Central agent 34 can also monitor the number of blocks 38 that are being used by port module 28 using a counter. When port module 28 notifies central agent 34 that port module 28 has written to a block 38, central agent increments the counter by an amount and, when a block 38 to which port module 28 has written is released and a credit corresponding to block 38 is returned to the pool of credits, central agent decrements the counter by an amount.

[0037] The number of credits that are available to a port module 28 can be kept constant, and the number of blocks 38 that are being used by port module 28 can be limited. The limit can be changed in response to changes in load conditions at port module 28, one or more other port module 28, or both. In particular embodiments, the number of blocks 38 that are being used by a port module 28 is limited according to a dynamic threshold that is a function of the number of credits in the pool of credits. An active port module 28, in particular embodiments, includes a port module 28 that is using one or more blocks 38. Reference to a port module 28 that is using a block 38 includes a port module 28 that has written at least one packet to stream memory 30 that has not been read from stream memory 30 to all designated port modules 28 of the packet. A dynamic threshold can include a fraction of the number of credits in the pool of credits calculated using the following formula, in which α equals the number of port modules 28 that are active and ρ is a parameter: 1$\frac{\rho }{1+\left(\rho \times \alpha \right)}$

[0038] A number of credits in the pool of credits can be reserved to prevent central agent 34 from allocating a credit to a port module 28 if the number of blocks 38 that are each being used by a port module 28 exceeds an applicable limit, which can include the dynamic threshold described above. Reserving one or more credits in the pool of credits can provide a cushion during a transient period associated with a change in the number of port modules 28 that are active. The fraction of credits that are reserved is calculated using the following formula, in which α equals the number of active port modules 28 and ρ is a parameter: 2$\frac{1}{1+\left(\rho \times \alpha \right)}$

[0039] According to the above formulas, if one port module 28 is active and ρ is two, central agent 34 reserves one third of the credits and may allocate up to two thirds of the credits to port module 28; if two port modules 28 are active and ρ is one, central agent 34 reserves one third of the credits and may allocate up to one third of the credits to each port module 28 that is active; and if twelve port modules 28 are active and ρ is 0.5, central agent 34 reserves two fourteenths of the credits and may allocate up to one fourteenth of the credits to each port module 28 that is active. Although a particular limit is described as being applied to the number of blocks 38 that are being used by a port module 28, the present invention contemplates any suitable limit being applied to the number of blocks 38 that are being used by a port module 28.

[0040] When a first port module 28 writes a packet to stream memory 30, first port module 28 can communicate to routing module 36 information from the header of the packet (such as one or more destination addresses) that routing module 36 can use to identify one or more second port modules 28 that are designated port modules 28 of the packet. First port module 28 can also communicate to routing module 36 an address of a first block 38 to which the packet has been written and an offset that together can be used by second port modules 28 to read the packet from stream memory 30. Routing module 36 can identify second port modules 28 using one or more routing tables and the information from the header of the packet and, after identifying second port modules 28, communicate the address of first block 38 and the offset to each second port module 28, which second port module 28 can add to an output queue, as described more fully below.

[0041] Central agent 34 returns a credit to the pool of credits only if all designated port modules 28 of a packet that has been written to a block 38 corresponding to the credit have read from block 38. As an example, consider a packet that has been written to a block 38 and that has two designated port modules 28. First designated port module 28 reads from block 38 and notifies central agent 34 that first designated port module 28 has read from block 38. Because second port module 28 has not yet read from block 38 and notified central agent 34 that second designated port module 28 has read from block 38, central agent 34 does not return a credit corresponding to block 38 to the pool of credits in response to the notification from first port module 28. Later, second designated port module 28 reads from block 38 and notifies central agent 34 that second designated port module 28 has read from block 38. Because first port module 28 has already read from block 38 and notified central agent 34 that first designated port module 28 has read from block 38, central agent 34 returns the credit corresponding to block 38 to the pool of credits in response to the notification from second port module 28.

[0042] To determine whether all designated port modules 28 of a packet have read from a block 38 to which the packet has been written, central agent 34 can use a bit vector. A bit vector can include two or more elements that each correspond to a port module 28 and indicate whether port module 28 has read from a block 38. When a packet is written to stream memory 30, central agent 34 can set the elements of a bit vector to indicate which port modules 28 of switch core 26 are designated port modules 28 of the packet, and, as designated port modules 28 read the packet from stream memory 30, central agent 34 can clear the elements of the bit vector.

[0043] As an example, consider a bit vector that includes six elements. A first element corresponds to a first port module 28, a second element corresponds to a second port module 28, a third element corresponds to a third port module 28, a fourth element corresponds to a fourth port module 28, a fifth element corresponds to a fifth port module 28, and a sixth element corresponds to a sixth port module 28. A packet is written to a block 38 of stream memory 30, and third port module 28, fourth port module 28, and sixth port module 28 are all designated port modules 28 of the packet. A third element of the bit vector corresponding to third port module 28 is set to indicate that third port module 28 is a designated port module 28 of the packet, a fourth element of the bit vector corresponding to fourth port module 28 is set to indicate that fourth port module 28 is a designated port module 28 of the packet, and a sixth element of the bit vector corresponding to sixth port module 28 is set to indicate that sixth port module 28 is a designated port module 28 of the packet. A first element of the bit vector, a second element of the bit vector, and a fifth element of the bit vector are all left clear, indicating that a first port module, a second port module 28, and a fifth port module 28, respectively, are not designated port modules 28.

[0044] Third port module 28 reads from block 38 first, and, when third port module 28 reads from block 38, the third element of the bit vector is cleared. The bit vector indicates that fourth port module 28 and sixth port module 28 have not yet read packet from block 38. Sixth port module 28 reads from block 38 next, and, when sixth port module 28 reads from block 38, the sixth element of the bit vector is cleared. The bit vector indicates that fourth port module 28 has not yet read from block 38. Fourth port module 28 reads from block 38 last, and, when fourth port module 28 reads from block 38, because fourth port module 28 is a last designated port module 28 to read from block 38, a credit corresponding to block 38 is returned to the pool of credits.

[0045] A bit vector can be stored in an entry of a multicast state table. The multicast state table can include multiple entries, at least some of which each correspond to a block 38 of stream memory 30. Each block 38 of stream memory 30 has a corresponding entry in tag memory 32. An error detection code (EDC) for detecting single- and multiple-bit errors can also be stored along with a bit vector in an entry in the multicast state table. When a packet has been written to stream memory 30, elements of a bit vector in an entry in the multicast state table corresponding to a first block 38 to which the packet has been written are set to indicate which port modules 28 are designated port modules 28 of the packet, as described above. Only the elements of the bit vector in the entry corresponding to first block 38 to which the packet has been written are set. When a designated port module 28 reads from first block 38, an element corresponding to designated port module 28 is cleared to indicate that designated port module 28 has started reading the packet from stream memory 30. When a last designated port module 28 reads from first block 38, central agent 34 returns a credit corresponding to first block 38 to the pool of credits. Central agent 34 returns credits corresponding to subsequent blocks 38 to which the packet has been written to the pool of credits as last designated port module 28 reads from subsequent blocks 38.

[0046] As an example, consider a packet that has been written to stream memory 30. A first portion of the packet has been written to a first block 38, and a second and final portion of the packet has been written to a second block 38. A first credit corresponds to first block 38, and a second credit corresponds to second block 38. A fifth port module 28 and a seventh port module 28 of switch core 26 are designated port modules 28 of the packet. A first entry in a multicast state table corresponds to first block 38, and second entry in the multicast state table corresponds to second block 38. Central agent 34 sets a fifth element and a seventh element of a bit vector in the first entry to indicate that fifth port module 28 and seventh port module 28, respectively, are designated port modules 28 of the packet. Central agent 34 need not set any elements of a bit vector in the second entry. Seventh port module 28 reads from first block 38 and notifies central agent 34 that seventh port module 28 has read from first block 38. Central agent 34 determines, from the bit vector in the first entry, that seventh port module 28 is not a last designated port module 28 to start reading the packet from stream memory 30 and clears the seventh element of the bit vector in the first entry, indicating that seventh port module 28 has started reading the packet from stream memory 30. Because seventh port module 28 is not a last designated port module 28 to start reading the packet from stream memory 30, central agent does not yet return the first credit to the pool of credits.

[0047] Fifth port module 28 reads from first port module 28 next and notifies central agent 34 that fifth port module 28 has read from first block 38. Central agent 34 determines, from the bit vector in the first entry, that fifth port module 28 is a last designated port module 28 to start reading the packet from stream memory 30 and, because fifth port module 28 is a last designated port module 28 to start reading the packet from stream memory 30, returns the first credit to the pool of credits. Seventh port module 28 then reads from second port module 28 and notifies central agent 34 that seventh port module 28 has read from second port module 28. Central agent determines, from the bit vector in the first entry, that seventh port module 28 is not a last designated port module 28 to start reading the packet from stream memory 30 and, because seventh port module 28 is not a last designated port module 28 to start reading the packet from stream memory 30, does not yet return second credit to the pool of credits. Fifth port module 28 reads from second port module 28 next and notifies central agent 34 that fifth port module 28 has read from second block 38. Central agent 34 determines, from the bit vector in the first entry, that fifth port module 28 is a last designated port module 28 to start reading the packet from stream memory 30 and, because fifth port module 28 is a last designated port module 28 to start reading the packet from stream memory 30, returns the second credit to the pool of credits.

[0048] In the above example, if fifth port module 28 overtook seventh port module 28 and read from second block 38 before seventh port module 28 read from second block 38, the second credit would be returned to the pool of credits before seventh port module 28 read from second block 38. To reduce the likelihood that fifth port module 28 will overtake seventh port module 28, fifth port module 28 and seventh port module 28 can both read from first block 38 and second block 38 at approximately the same speed.

[0049] Also, in the above example, if the first credit, after being returned to the pool of credits, were allocated to a port module 28 and used to write to first block 38 before second port module 28 had read from second block 38, the bit vector in the first entry would be overwritten such that central agent 34 would be unable to determine whether fifth port module 28 or seventh port module 28 were a last port module 28 to start reading the packet from stream memory 30. To reduce the likelihood that the bit vector in the first entry will be overwritten, a dynamic threshold can be applied to the number of credits that are available to a port module 28, as described above. The dynamic threshold can prevent the number of credits in the pool of credits from becoming so small that all designated port modules 28 of a packet do not have enough time to read the packet from stream memory 30 before a bit vector is overwritten in an entry in the multicast state table corresponding to a first block 38 to which the packet has been written.

[0050] A port module 28 can include one or more output queues that are used to queue packets that have been written to stream memory 30 for communication out of switch core 26 through port module 28. When a packet is written to stream memory 30, the packet is added to an output queue of each designated port module 28 of the packet. An output queue of a designated port module 28 can correspond to a combination of a level of quality of service (QoS) and a source port module 28. As an example, consider a switch core 26 that provides three levels of QoS and includes four port modules 28 including both input logic and output logic. A first port module 28 includes nine output queues: a first output queue corresponding to the first level of QoS and a second port module 28; a second output queue corresponding to the first level of QoS and a third port module 28; a third output queue corresponding to the first level of QoS and a fourth port module 28; a fourth output queue corresponding to the second level of QoS and second port module 28; a fifth output queue corresponding to the second level of QoS and third port module 28; a sixth output queue corresponding to the second level of QoS and fourth port module 28; a seventh output queue corresponding to the third level of QoS and second port module 28; an eighth output queue corresponding to the third level of QoS and third port module 28; and a ninth output queue corresponding to the third level of QoS and fourth port module 28. A packet that has been written to stream memory 30 is added to the first output queue of first port module 28 if (1) the packet has been written to stream memory 30 from second port module 28, (2) first port module 28 is a designated port module 28 of the packet, and (3) the level of QoS of the packet is the first level of QoS. A packet that has been written to stream memory 30 is added to the fifth output queue of first port module 28 if (1) the packet has been written to stream memory 30 from third port module 28, (2) first port module 28 is a designated port module 28 of the packet, and (3) the level of QoS of the packet is the second level of QoS. A packet that has been written to stream memory 30 is added to the ninth output queue of first port module 28 if (1) the packet has been written to stream memory 30 from fourth port module 28, (2) first port module 28 is a designated port module 28 of the packet, and (3) the level of QoS of the packet is the third level of QoS.

[0051] Second port module 28 also includes nine output queues: a first output queue corresponding to the first level of QoS and a first port module 28; a second output queue corresponding to the first level of QoS and a third port module 28; a third output queue corresponding to the first level of QoS and a fourth port module 28; a fourth output queue corresponding to the second level of QoS and first port module 28; a fifth output queue corresponding to the second level of QoS and third port module 28; a sixth output queue corresponding to the second level of QoS and fourth port module 28; a seventh output queue corresponding to the third level of QoS and first port module 28; an eighth output queue corresponding to the third level of QoS and third port module 28; and a ninth output queue corresponding to the third level of QoS and fourth port module 28. A packet that has been written to stream memory 30 is added to the first output queue of second port module 28 if (1) the packet has been written to stream memory 30 from first port module 28, (2) second port module 28 is a designated port module 28 of the packet, and (3) the level of QoS of the packet is the first level of QoS. A packet that has been written to stream memory 30 is added to the fifth output queue of second port module 28 if (1) the packet has been written to stream memory 30 from third port module 28, (2) second port module 28 is a designated port module 28 of the packet, and (3) the level of QoS of the packet is the second level of QoS. A packet that has been written to stream memory 30 is added to the ninth output queue of second port module 28 if (1) the packet has been written to stream memory 30 from fourth port module 28, (2) second port module 28 is a designated port module 28 of the packet, and (3) the level of QoS of the packet is the third level of QoS.

[0052] Third port module 28 and fourth port module 28 each include output queues similar to the output queues of first port module 28 and the output queues of second port module 28 described above. QoS can encompass rate of transmission, rate of error, or other aspect of the communication of packets through switch core 26, and reference to QoS can include class of service (CoS), where appropriate. Although an output queue of a first port module 28 is described as corresponding to a second port module 28 and a level of QoS, an output queue of a first port module 28 need not necessarily correspond to a second port module 28 and a level of QoS. As an example, in particular embodiments, an output queue of a first port module 28 can correspond to a second port module 28 and not a level of QoS.

[0053] An output queue of a port module 28 includes a register of port module 28 and, if there is more than one packet in the output queue, one or more entries in a memory structure of port module 28, as described below. A port module 28 includes a memory structure that can include one or more linked lists that port module 28 can use, along with one or more registers, to determine a next packet to read from stream memory 30. The memory structure includes multiple entries, at least some of which each correspond to a block 38 of stream memory 30. Each block 38 of stream memory 30 has a corresponding entry in the memory structure. An entry in the memory structure can include a pointer to another entry in the memory structure, resulting in a linked list. A port module 28 also includes one or more registers that port module 28 can also use to determine a next packet to read from stream memory 30. A register includes a write pointer, an offset, and a read pointer. The write pointer can point to a first block 38 to which a first packet has been written, the offset can indicate a first word 40 to which the first packet has been written, and the read pointer can point to a first block 38 to which a second packet (which could be the same packet as or a packet other than the first packet) has been written. Because entries in the memory structure each correspond to a block 38 of stream memory 30, a pointer that points to a block 38 also points to an entry in the memory structure.

[0054] Port module 28 can use the write pointer to determine a next entry in the memory structure to which to write an offset. Port module 28 can use the offset to determine a word 40 of a block 38 at which to start reading from block 38. Port module 28 can use the read pointer to determine a next packet to read from stream memory 30. Port module 28 can also use the write pointer and the read pointer to determine whether more than one packet is in the output queue. If output queue is not empty and the write pointer and the read pointer both point to the same block 38, there is only one packet in the output queue. If there is only one packet in the output queue, port module 28 can determine a next packet to read from stream memory 30 and read the next packet from stream memory 30 without accessing the memory structure.

[0055] If a first packet is added to the output queue when there are no packets in the output queue, (1) the write pointer in the register is modified to point to a first block 38 to which the first packet has been written, (2) the offset is modified to indicate a first word 40 to which the first packet has been written, and (3) the read pointer is also modified to point to first block 38 to which the first packet has been written. If a second packet is added to the output queue before port module 28 reads the first packet from stream memory 30, (1) the write pointer is modified to point to a first block 38 to which the second packet has been written, (2) the offset is written to a first entry in the memory structure corresponding to first block 38 to which the first packet has been written and then modified to indicate a first word 40 to which the second packet has been written, and (3) a pointer in the first entry is modified to point to first block 38 to which the second packet has been written. The read pointer is left unchanged such that, after the second packet is added to the output queue, the read pointer still points to first block 38 to which the first packet has been written. As described more fully below, the read pointer is changed when port module 28 reads a packet in the output queue from stream memory 30. If a third packet is added to the output queue before port module 28 reads the first packet and the second packet from stream memory 30, (1) the write pointer is modified to point to a first block 38 to which the third packet has been written, (2) the offset is written to a second entry in the memory structure corresponding to first block 38 to which the second packet has been written and modified to indicate a first word 40 to which the third packet has been written, and (3) a pointer in the second entry is modified to point to first block 38 to which the third packet has been written. The read pointer is again left unchanged such that, after the third packet is added to the output queue, the read pointer still points to first block 38 to which the first packet has been written.

[0056] Port module 28 can use the output queue to determine a next packet to read from stream memory 30. As an example, consider the output queue described above in which there are three packets. In the register, (1) the write pointer points to first block 38 to which the third packet has been written, (2) the offset indicates first word 40 to which the third packet has been written, and (3) the read pointer points to first block 38 to which the first packet has been written. The first entry in the memory structure includes (1) an offset that indicates first word 40 to which the first packet has been written and (2) a pointer that points to first block 38 to which the second packet has been written. The second entry in the memory structure includes (1) an offset that indicates first word 40 to which the second packet has been written and (2) a pointer that points to first block 38 to which the third packet has been written.

[0057] Port module 28 compares the read pointer with the write pointer and determines, from the comparison, that there is more than one packet in the output queue. Port module 28 then uses the read pointer to determine a next packet to read from stream memory 30. The read pointer refers port module 28 to first block 38 of the first packet, and, since there is more than one packet in the output queue, port module 28 accesses the offset in the first entry indicating first word 40 to which the first packet has been written. Port module 28 then reads the first packet from stream memory 30, using the offset in the first entry, starting at first block 38 to which the first packet has been written. If the first packet has been written to more than one block 38, port module 28 can use a linked list in tag memory 32 to read the first packet from memory, as described above.

[0058] While port module 28 is reading the first packet from stream memory 30, port module 28 copies the pointer in the first entry to the read pointer, compares the read pointer with the write pointer, and determines, from the comparison, that there is more than one packet in the output queue. Port module 28 then uses the read pointer to determine a next packet to read from stream memory 30. The read pointer refers port module 28 to first block 38 of the second packet, and, since there is more than one packet in the output queue, port module 28 accesses the offset in the second entry indicating first word 40 to which the second packet has been written. When port module 28 has finished reading the first packet from stream memory 30, port module 28 reads the second packet from stream memory 30, using the offset in the second entry, starting at first block 38 to which the second packet has been written. If the second packet has been written to more than one block 38, port module 28 can use a linked list in tag memory 32 to read the second packet from memory, as described above.

[0059] While port module 28 is reading the second packet from stream memory 30, port module 28 copies the pointer in the second entry to the read pointer, compares the read pointer with the write pointer, and determines, from the comparison, that there is only one packet in the output queue. Port module 28 then uses the read pointer to determine a next packet to read from stream memory 30. The read pointer refers port module 28 to third block 38 of the second packet, and, since there is only one packet in the output queue, port module 28 accesses the offset in the register indicating first word 40 to which the third packet has been written. When port module 28 has finished reading the second packet from stream memory 30, port module 28 reads the third packet from stream memory 30, using the offset in the register, starting at first block 38 to which the third packet has been written. If the third packet has been written to more than one block 38, port module 28 can use a linked list in tag memory 32 to read the third packet from memory, as described above.

[0060] If a port module 28 includes more than one output queue, an algorithm can be used for arbitration among the output queues. Arbitration among multiple output queues can include determining a next output queue to use to determine a next packet to read from stream memory 30. Arbitration among multiple output queues can also include determining how many packets in a first output queue to read from stream memory 30 before using a second output queue to determine a next packet to read from stream memory 30. The present invention contemplates any suitable algorithm for arbitration among multiple output queues. As an example and not by way of limitation, according to an algorithm for arbitration among multiple output queues of a port module 28, port module 28 accesses output queues that are not empty in a series of rounds. In a round, port module 28 successively accesses the output queues in a predetermined order and, when port module 28 accesses an output queue, reads one or more packets in the output queue from stream memory 30. The number of packets that port module 28 reads from an output queue in a round can be the same as or different from the number of packets that port module 28 reads from each of one or more other output queues of port module 28 in the same round. In particular embodiments, the number of packets that can be read from an output queue in a round is based on a quantum value that defines an amount of data according to which more packets can be read form the output queue if smaller packets are in the output queue and fewer packets can be read from the output queue if larger packets are in the output queue, which can facilitate fair sharing of an output link of port module 28.

[0061] In particular embodiments, a port module 28 uses a connection to access stream memory 30. In these embodiments, port module 28 establishes a connection to stream memory 30, accesses stream memory 30 using the connection, and, if necessary, releases the connection. When accessing stream memory 30 using a connection, port module 28 experiences no blocking by other port modules 28. In particular embodiments, there is always a connection between a port module 28 and stream memory 30 (and there is no arbitration delay) for write operations. A write operation includes a number of steps over a series of cycles (each of which includes one or more clock cycles of switch core 26). As an example and not by way of limitation, stream memory 30 communicates one or more sync bits to port module 28 (which indicate a word offset for the write operation) and port module 28 writes to stream memory 30 and communicates one or more addresses of one or more blocks 38 of stream memory 30 for the write operation to stream memory 30.

[0062] A read operation (in which arbitration and access are pipelined) also includes a number of steps over a series of cycles. As an example and not by way of limitation, port module 28 requests a connection for a read operation from stream memory 30 and communicates a word offset to stream memory 30. After an arbitration cycle spanning one or more clock cycles of switch core 26, stream memory 30 communicates an acknowledgement to port module 28 in response to the request, at which point the requested connection is established. In particular embodiments, there is an estimated minimum arbitration delay (which includes a delay between a connection being requested and an acknowledgement being communicated) of zero clock cycles and a maximum estimated arbitration delay of fourteen clock cycles. Arbitration delay causes gaps in streams of data through switch core 26, and the average arbitration delay that port module 28 experiences tends to increase as the load experienced by switch core 26 increases. After port module 28 receives the acknowledgement, port module 28 communicates to stream memory 30 one or more addresses of blocks 38 of stream memory 30 for the read operation, and, one or more cycles later, stream memory 30 communicates the data at those addresses to port module 28. Stream memory 30 can begin to communicate the data before port module 28 has communicated to stream memory 30 all the addresses for the read operation. When port module 28 has communicated to stream memory 30 all the addresses for the read operation, port module 28 releases the connection and, in particular embodiments, requests another connection, possibly in the same cycle. More read operations from stream memory 30 than write operations to stream memory 30 can be scheduled over a period of time. As an example and not by way of limitation, twice as many read operations can be scheduled over a period of time than write operations over the same period of time. As another example, three times as many read operations can be scheduled over a period of time than write operations over the same period of time.

[0063] In particular embodiments, stream memory 30 includes a number of static random access memory (SRAM) devices used in parallel with each other, and access to the SRAM devices of stream memory 30 is scheduled using an appropriate interleaving technique. The present invention contemplates 1RW (or single port) SRAM devices, multi-port and multi-bit SRAM devices, or other SRAM devices, although 1RW SRAM devices provide greater density, flexibility, and fewer wires for access to streams of data. If switch core 26 includes N port modules 28 and the links between stream memory 30 and port modules 38 each carry M bits, stream memory 30, in particular embodiments, includes 2*N instances of 1RW SRAM devices having data paths that are 2*M bits wide. As an example and not by way of limitation, if switch core 26 includes twelve port modules 28 and the links coupling port modules 28 to stream memory 30 each carry thirty-six bits, stream memory 30 includes twenty-four instances of 1RW SRAM devices having data paths that are seventy-two bits wide. The number of instances of SRAM devices and the width of the data paths are based on the following observations: (1) the total bandwidth of port modules 28 is N*M bits per second for read operations and N*M bits per second for write operations, and the total bandwidth of stream memory 30 is 4*N*M bits per second; (2) read operations and write operations to and from stream memory 30 are scheduled such that N*M bits per second are reserved for write operations and 3*N*M bits per second are reserved for read operations; and (3) providing two to three times more bandwidth for read operations than for write operations reduces the arbitration delay of switch core 26. Although stream memory 30 is described as including SRAM devices, the present invention contemplates stream memory 30 including any suitable memory devices.

[0064] In particular embodiments, a multistage interconnection network (MIN) including switching structure 34a and switching structure 34b is used to provide connections between all the SRAM devices of stream memory 30 and all port modules 28 of switch core. The MIN includes a hierarchical structure including a number of switching units 42 and a number of memory banks 44 (into which the SRAM devices of stream memory 30 are organized) that include bank switching units 46, as described more fully below. The MIN of stream memory 30 illustrated in FIG. 3 includes four switching units 42 and three memory banks 44. Although stream memory 30 is described and illustrated as including a particular number of switching units 42 and a particular number of memory banks 44 in a particular configuration, the present invention contemplates stream memory 30 including any suitable number of switching units 42 and any suitable number of memory banks 44 in any suitable configuration. Bank switching units 46 include statically scheduled, regular switching units. In particular embodiments, static scheduling is used for write operations and on-demand scheduling at switching units 42 is used for read operations. The MIN of stream memory 30 is nonblocking, but without redundancy.

[0065] A switching unit 42 can receive all or a portion of a packet from a port module 28 and switch the received data to a memory bank 44. Write operations via a switching unit 42 are scheduled according to any suitable technique. As an example, static scheduling at a switching unit 42 is used for write operations. FIG. 5 illustrates example scheduling at two switching units 42 of switch core 26 for write operations to three memory banks 44 by six port modules 28. In particular embodiments, as an example and not by way of limitation, switching units 42 return to an initial state every forty-eight cycles. Over the forty-eight cycles, each port module 28 is given an opportunity to write to each memory unit 48 (which are described more fully below) of each memory bank 44. Although switching units 42 are described as returning to an initial state every forty-eight cycles, the present invention contemplates switching units 42 returning to an initial state after any suitable number of cycles. Each switching unit 42 has three states and changes states every sixteen cycles. Although a particular schedule at a particular number of switching units 42 over a particular number of cycles is described and illustrated for write operations to a particular number of memory banks 44 by a particular number of port modules 28, the present invention contemplates any suitable schedule at any suitable number of switching units 42 over any suitable number of cycles for write operations to any suitable number of memory banks 44 by any suitable number of port modules 28.

[0066] Switching unit 42 can also receive all or a portion of a packet from a memory bank 44 and switches the received data to a port module 28. Read operations via a switching unit 42 are scheduled according to any suitable technique. As an example, on-demand scheduling at a switching unit 42 is used for read operations. This scheduling includes a connect and release technique, since more than one port module 28 could attempt to read from a memory unit 48 in the same cycle. If static scheduling were used, in particular embodiments, a port module 28 would have to wait up to forty-eight cycles to read from a particular memory unit 48 of a particular memory bank 44. To reduce this delay, arbitration at a switching unit 42 among port modules 28 coupled to switching unit 42 is used for read operations. The availability of each memory unit 48 for read operations is monitored and a particular port module 28 is allowed to read from a particular memory unit 48 of a particular memory bank 44 every four or eight cycles unless another port module 28 is reading from memory unit 48.

[0067] FIG. 6 illustrates example scheduling at a switching unit 42 of switch core 26 for read operations from twenty-four memory units 48 (within memory banks 44) by a port module 28. Three port modules 28 are coupled to switching unit 42, and each memory unit 48 is designated in schedule 50 by a number from zero to twenty-three. Read operations span two cycles and, according to schedule 50, begin in cycles zero, two, four, six, eight, ten, twelve, and fourteen. Port module 28 can read from any one of nine memory units 48 in a read cycle (which include two cycles for a read operation). In the read cycle spanning cycles zero and one, port module 28 can read from memory unit 48 designated 2, 4, 6, 10, 12, 14, 18, 20, or 22 if no other port module 28 coupled to switching unit 42 is reading from memory unit 48; in the read cycle spanning cycles two and three, port module 28 can read from memory unit 48 designated 3, 5, 7, 11, 13, 15, 19, 21, or 23 if no other port module 28 coupled to switching unit 42 is reading from memory unit 48; in the read cycle spanning cycles four and five, port module 28 can read from memory unit 48 designated 0, 4, 6, 8, 12, 14, 16, 20, or 22 if no other port module 28 coupled to switching unit 42 is reading from memory unit 48; and so on. Similar scheduling at switching unit 42 is used for read operations from memory units 38 to other port modules 28 coupled to switching unit 42 and to other port modules 28 coupled to other switching units in switch core 26. Although a particular schedule at a switching unit 42 over a particular number of cycles is described and illustrated for read operations from a particular number of memory units 48 by a port module 28, the present invention contemplates any suitable schedule at a switching unit 42 over any suitable number of cycles for read operations from any suitable number of memory units 48 by a port module 28.

[0068] FIG. 7 illustrates an example memory bank 44 of switch core 26. A memory bank 44 includes one or more memory units 48 and one or more bank switching units 46 (which include statically scheduled, regular switching units). In particular embodiments, one or more memory units 48 together include a memory structure 36. In particular embodiments, memory bank 44 includes built-in self-test (BIST) logic. Memory bank 44 is shared by port modules 28, which, in particular embodiments, eliminates head-of-line blocking (thereby increasing the throughput of switch core 26), enables switch core 26 to more efficiently handle changes in load conditions at port modules 28, and reduces memory requirements associated with switch core 26. In particular embodiments, as an example and not by way of limitation, memory bank 44 includes eighteen bank switching units 46 and eight memory units 48 (as illustrated in FIG. 7). Although memory bank 44 is described and illustrated as including a particular number of bank switching units 46 and a particular number of memory units 48, the present invention contemplates memory bank 44 including any suitable number of bank switching units 46 and any suitable number of memory units 48 in any suitable configuration. A memory unit 48 includes one or more SRAM devices. As an example and not by way of limitation, if stream memory 30 includes twenty-four instances of SRAM devices (as described above), stream memory 30 includes three memory banks 44, each memory bank 44 includes eight memory units 48, and each memory unit 48 includes one SRAM device. In particular embodiments, if stream memory 30 is logically divided into 1536 blocks 38 and includes twenty-four memory units 48, each memory unit 48 includes sixty-four blocks 38 of stream memory 30.

[0069] The link coupling memory bank 44 to a switching unit 42 includes one or more links. As an example, in particular embodiments, the link coupling memory bank 44 to switching unit 42 includes five links, one for write operations and four for read operations, and each of these links carries thirty-six bits. Memory bank 44 illustrated in FIG. 7 is coupled to four switching units 42, designated ul (up left), dl (down left), ur (up right), and dr (down right), respectively. As an example, ul designates switching unit 42a, dl designates switching unit 42b, ur designates switching unit 42c, and dr designates switching unit 42d. Links designated W are each for write operations to any memory unit 36 from a switching unit 42, and links designated R are each for read operations from particular memory units 48 to switching unit 42. Specifically, links designated R01 are for read operations from memory unit 48a and memory unit 48b; links designated R23 are for read operations from memory unit 48c and memory unit 48d; links designated R45 are for read operations from memory unit 48e and memory unit 48f; and links designated R67 are for read operations from memory unit 48g and memory unit 48h. The links designated -ul couple memory bank 44 to switching unit 42a, the links designated -dl couple memory bank 44 to switching unit 42b, the links designated -ur couple memory bank 44 to switching unit 42c, and the links designated -dr couple memory bank 44 to switching unit 42d. Thus, the link designated R01-ul is for read operations from memory units 48a and 46b to switching unit 42a, the link designated W-dr is for write operations to any memory unit 48 from switching unit 42d, the link designate R67-ur is for read operations from memory units 48g and 46h to switching unit 42c, and so on. If switching unit 42 is coupled to three port modules 28 and three memory banks 44 (as illustrated in FIG. 3), switching unit 42 includes a 3×3, 36-bit switching unit for write operations and a 12×3, 36-bit switching unit for read operations.

[0070] As described above, bank switching units 46 include statically scheduled, regular switching units. FIG. 8 illustrates example scheduling at three bank switching units 46 of a memory bank 44 for read operations to two memory units 48 via four switching units 42. In cycle N, the link designated R01-ul (which is for read operations from memory units 48a and 46b to switching unit 42a) is scheduled to read from memory unit 48a, and the link designated R01-ur (which is for read operations from memory units 48a and 46b to switching unit 42c) is scheduled to read from memory unit 48b. In cycle N+1, the link designated R01-dl (which is for read operations from memory units 48a and 46b to switching unit 42b) is scheduled to read from memory unit 48a, and the link designated R01-dr (which is for read operations from memory units 48a and 46b to switching unit 42d) is scheduled to read from memory unit 48b. In cycle N+2, the link designated R01-ul is scheduled to read from memory unit 48b, and the link designated R01-ur is scheduled to read from memory unit 48a. And, in cycle N+3, the link designated R01-dl is scheduled to read from memory unit 48b, and the link designated R01-dr is scheduled to read from memory unit 48a. Similar scheduling is used for read operations to other pairs of memory units 48, such as memory units 48c and 46d, memory units 48e and 46f, and memory units 48g and 46h. Although a particular schedule at a particular number of bank switching units 46 over a particular number of cycles for read operations to a particular number of memory units 48 via a particular number of switching units 42 is described and illustrated, the present invention contemplates any suitable schedule at any suitable number of bank switching units 46 over any suitable number of cycles for read operations to any suitable number of memory units 48 via any suitable number of switching units 42.

[0071] FIG. 9 illustrates example scheduling for write operations to and read operations from eight memory units 48 of a memory bank 44 via four switching units 42. Each memory unit 48 is designated in schedule 52 by a number from zero to seven: memory unit 48a is designated by the number zero; memory unit 48b is designated by the number one; memory unit 48c is designated by the number two; memory unit 48d is designated by the number three; memory unit 48e is designated by the number four; memory unit 48f is designated by the number five; memory unit 48g is designated by the number six; and memory unit 48h is designated by the number seven. Upper half 54 of schedule 52 applies to the links coupling switching units 42a and 40b to memory bank 44, and lower half 56 of schedule 52 applies to the links coupling switching units 42c and 40d to memory bank 44. Columns 58 corresponding to even cycles (zero, two, four, six, eight, ten, twelve, and fourteen) apply to the links coupling switching units 42a and 40c to memory bank 44, and columns 58 corresponding to odd cycles (one, three, five, seven, nine, eleven, thirteen, and fifteen) apply to the links coupling switching units 42b and 40d to memory bank 44. Rows 60 correspond, respectively, to the links coupling switching units 42a, 40b, 40c, and 40d to memory bank 44. Areas 62 indicate where read operations cannot take place due to conflicts with write operations.

[0072] According to schedule 52, at cycle zero, the link designated W-ul (which is for write operations via switching unit 42a) can be used for one or more write operations to memory unit 48a; the link designated R23-ul (which is for read operations from memory units 48c and 46d via switching unit 42a) can be used for one or more read operations from memory unit 48c; the link designated R45-ul (which is for read operations from memory units 48e and 46f via switching unit 42a) can be used for one or more read operations from memory unit 48e; the link designated R67-ul (which is for read operations from memory units 48g and 46h via switching unit 42a) can be used for one or more read operations from memory unit 48g; the link designated W-ur (which is for write operations via switching unit 42c) can be used for one or more write operations to memory unit 48d; the link designated R01-ur (which is for read operations from memory units 48c and 46d via switching unit 42c) can be used for one or more read operations from memory unit 48b; the link designated R45-ur (which is for read operations from memory units 48e and 46f via switching unit 42c) can be used for one or more read operations from memory unit 48f, and the link designated R67-ur (which is for read operations from memory units 48g and 46h via switching unit 42c) can be used for one or more read operations from memory unit 48h.

[0073] At cycle one, the link designated W-dl (which is for write operations via switching unit 42b) can be used for one or more write operations to memory unit 48a; the link designated R23-dl (which is for read operations from memory units 48c and 46d via switching unit 42b) can be used for one or more read operations from memory unit 48c; the link designated R45-dl (which is for read operations from memory units 48e and 46f via switching unit 42b) can be used for one or more read operations from memory unit 48e; the link designated R67-dl (which is for read operations from memory units 48g and 46h via switching unit 42b) can be used for one or more read operations from memory unit 48g; the link designated W-dr (which is for write operations via switching unit 42d) can be used for one or more write operations to memory unit 48d; the link designated R01-dr (which is for read operations from memory units 48c and 46d via switching unit 42d) can be used for one or more read operations from memory unit 48b; the link designated R45-dr (which is for read operations from memory units 48e and 46f via switching unit 42d) can be used for one or more read operations from memory unit 48f; and the link designated R67-dr (which is for read operations from memory units 48g and 46h via switching unit 42d) can be used for one or more read operations from memory unit 48h.

[0074] At cycle two, the link designated W-ul can be used for one or more write operations to memory unit 48b; the link designated R23-ul can be used for one or more read operations from memory unit 48d; the link designated R45-ul can be used for one or more read operations from memory unit 48f, the link designated R67-ul can be used for one or more read operations from memory unit 48h; the link designated W-ur can be used for one or more write operations to memory unit 48e; the link designated RO-ur can be used for one or more read operations from memory unit 48a; the link designated R23-ur (which is for read operations from memory units 48c and 46d via switching unit 42c) can be used for one or more read operations from memory unit 48c; and the link designated R67-ur can be used for one or more read operations from memory unit 48g.

[0075] This process continues according to schedule 52, reaching the initial state (which is cycle zero) after sixteen cycles. Although a particular schedule for write operations to and read operation from a particular number of memory units 48 of a memory bank 44 via a particular number of switching units 42 over a particular number of cycles is described and illustrated, the present invention contemplates any suitable schedule for write operations to and read operation from any suitable number of memory units 48 of a memory bank 44 via any suitable number of switching units 42 over any suitable number of cycles.

[0076] FIG. 10 illustrates an example method for memory interleaving in a high-speed switching environment. The method begins at step 100, where a first port module 28 of a switch core 26 receives a portion of a packet from a port 24 coupled to first port module 28. As described above, switch core 26 includes multiple port modules 28. At step 102, first port module 28 communicates the received portion of the packet to stream memory 30. Stream memory 30 is shared by all port modules 28 of switch core 26 and is nonblocking. In particular embodiments, as described above, stream memory 30 includes a MIN that couples a number of memory units 48 to port modules 28. The MIN includes a hierarchical structure of switching units 42 and memory banks 42, each including a number of bank switching units 44 and a number of memory units 48. As described above, in particular embodiments, stream memory 30 includes four switching units 42 and three memory banks 42 that each include eight memory units 48. As described above, in particular embodiments, static scheduling is used for write operations to memory units 48 of stream memory 30 and on-demand scheduling at switching units 42 is used for read operations from memory units 48 of stream memory 30.

[0077] At step 104, the MIN of stream memory 30 receives the communicated portion of the packet and, using one or more scheduling techniques, switches the received portion of the packet to an appropriate memory unit 48 of stream memory 30. At step 106, the switched portion of the packet is written to memory unit 48. At step 108, a second port module 28 requests a connection to memory unit 48 to which the portion of the packet was written. At step 110, after an arbitration cycle, the request is granted and the requested connection is established. As described above, arbitration cycles can vary in length from one read operation to the next. At step 112, the portion of the packet written to memory unit 48 is communicated to second port module 28 via the MIN of stream memory 30 using the established connection. At step, 114 second port module 28 receives the communicated portion of the packet and communicates the received portion of the packet out of switch core 26 to a port 24 coupled to second port module 28, at which point the method ends. Although particular steps of the method illustrated in FIG. 10 are described and illustrated as occurring in a particular order, the present invention contemplates any suitable steps of the method described above occurring in any suitable order.

[0078] Although the present invention has been described with several embodiments, sundry changes, substitutions, variations, alterations, and modifications may be suggested to one skilled in the art, and it is intended that the invention may encompass all such changes, substitutions, variations, alterations, and modifications falling within the spirit and scope of the appended claims.

Claims

1. A system for memory interleaving in a high-speed switching environment, the system comprising: a plurality of memory units that each comprise one or more memory devices; a plurality of port modules that are each operable to: receive a packet communicated from a component of a communications network and write the received packet to one or more of the plurality of memory units; and read a packet from one or more of the plurality of memory units for communication to the component of the communications network; and an interconnection network comprising a hierarchical structure that comprises one or more switching stages, the interconnection network coupling the plurality of memory units to the plurality of port modules such that: each of the port modules is operable to write to each of the memory units according to a first schedule and each of the port modules is operable to read from each of the memory units according to a second schedule, the first schedule allowing a first number of write operations over a period of time, the second schedule allowing a second number of read operations over the period of time, the second number being twice or more the first number; and a first port module is operable to read a first portion of a packet from one or more memory units for communication to a first component of the communications network before a second port module has received a second portion of the packet communicated from a second component of the communications network.

2. The system of claim 1, wherein the interconnection network comprises a multistage interconnection network (MIN) comprising one or more memory banks and one or more switching units coupling the plurality of port modules to the one or more memory banks, the one or more memory banks each comprising one or more of the plurality of memory units and one or more bank switching units.

3. The system of claim 2, wherein: each port module is coupled to a switching unit by a first link for write operations and a second link for read operations; and each switching unit is coupled to each memory bank by a third link for write operations and four fourth links for read operations.

4. The system of claim 2, wherein the MIN comprises three memory banks and four switching units, each memory bank comprising eight memory units that each comprise one memory device.

5. The system of claim 1, wherein the memory devices comprise static random access memory (SRAM) devices.

6. The system of claim 5, wherein the SRAM devices comprise one port for read operations and one port for write operations.

7. The system of claim 1, wherein the system is embodied in a single integrated circuit (IC).

8. The system of claim 1, wherein the high-speed switching environment is an Ethernet switching environment, an INFINIBAND switching environment, a 3GIO switching environment, a HYPERTRANSPORT switching environment, a RAPID IO switching environment, or a proprietary backplane switching environment.

9. The system of claim 1, wherein the first schedule comprises a static schedule and the second schedule comprises an on-demand schedule.

10. The system of claim 1, wherein write operations and read operations to the plurality of memory units are nonblocking.

11. A method for memory interleaving in a high-speed switching environment, the method comprising: using an interconnection network comprising a hierarchical structure that comprises one or more switching stages, coupling a plurality of memory units to a plurality of port modules, each memory unit comprising one or more memory devices, each port module being operable to receive a packet communicated from a component of a communications network and write the received packet to one or more of the plurality of memory units, each memory unit being further operable to read a packet from one or more of the plurality of memory units for communication to the component of the communications network, the plurality of memory units being coupled to the plurality of port modules such that: each of the port modules is operable to write to each of the memory units according to a first schedule and each of the port modules is operable to read from each of the memory units according to a second schedule, the first schedule allowing a first number of write operations over a period of time, the second schedule allowing a second number of read operations over the period of time, the second number being twice or more the first number; and a first port module is operable to read a first portion of a packet from one or more memory units for communication to a first component of the communications network before a second port module has received a second portion of the packet communicated from a second component of the communications network.

12. The method of claim 11, wherein the interconnection network comprises a multistage interconnection network (MIN) comprising one or more memory banks and one or more switching units coupling the plurality of port modules to the one or more memory banks, the one or more memory banks each comprising one or more of the plurality of memory units and one or more bank switching units.

13. The method of claim 12, wherein: each port module is coupled to a switching unit by a first link for write operations and a second link for read operations; and each switching unit is coupled to each memory bank by a third link for write operations and four fourth links for read operations.

14. The method of claim 12, wherein the MIN comprises three memory banks and four switching units, each memory bank comprising eight memory units that each comprise one memory device.

15. The method of claim 11, wherein the memory devices comprise static random access memory (SRAM) devices.

16. The system of claim 15, wherein the SRAM devices comprise one port for read operations and one port for write operations.

17. The method of claim 11, wherein the system is embodied in a single integrated circuit (IC).

18. The method of claim 11, wherein the high-speed switching environment is an Ethernet switching environment, an INFINIBAND switching environment, a 3GIO switching environment, a HYPERTRANSPORT switching environment, a RAPID IO switching environment, or a proprietary backplane switching environment.

19. The method of claim 11, wherein the first schedule comprises a static schedule and the second schedule comprises an on-demand schedule.

20. The method of claim 11, wherein write operations and read operations to the plurality of memory units are nonblocking.

21. A system for memory interleaving in a high-speed switching environment that comprises an Ethernet switching environment, an INFINIBAND switching environment, a 3GIO switching environment, a HYPERTRANSPORT switching environment, a RAPID IO switching environment, or a proprietary backplane switching environment, the system being embodied in a single integrated circuit (IC) and comprising: twenty-four memory units that each comprise one memory device comprising one static random access memory (SRAM), the SRAM device comprising one port for read operations and one port for write operations; twelve port modules that are each operable to: receive a packet communicated from a component of a communications network and write the received packet to one or more of the twenty-four memory units; and read a packet from one or more of the twenty-four memory units for communication to the component of the communications network; and a multistage interconnection network (MIN) comprising a hierarchical structure that comprises at least two switching stages, the MIN comprising three memory banks and four switching units coupling the plurality of port modules to the three memory banks, the three memory banks each comprising eight of the twenty-four memory units and eighteen bank switching units, each port module being coupled to a switching unit by a first link for write operations and a second link for read operations, each switching unit being coupled to each memory bank by a third link for write operations and four fourth links for read operations, the MIN coupling the twenty-four memory units to the twelve port modules such that: each of the port modules is operable to write to each of the memory units according to a first schedule and each of the port modules is operable to read from each of the memory units according to a second schedule, the first schedule comprising a static schedule, the second schedule comprising an on-demand schedule, the first schedule allowing a first number of write operations over a period of time, the second schedule allowing a second number of read operations over the period of time, the second number being twice or more the first number; and a first port module is operable to read a first portion of a packet from one or more memory units for communication to a first component of the communications network before a second port module has received a second portion of the packet communicated from a second component of the communications network.

22. A method for memory interleaving in a high-speed switching environment, the method comprising: means for coupling a plurality of memory units to a plurality of port modules, each memory unit comprising one or more memory devices, each port module being operable to receive a packet communicated from a component of a communications network and write the received packet to one or more of the plurality of memory units, each memory unit being further operable to read a packet from one or more of the plurality of memory units for communication to the component of the communications network, the plurality of memory units being coupled to the plurality of port modules such that: each of the port modules is operable to write to each of the memory units according to a first schedule and each of the port modules is operable to read from each of the memory units according to a second schedule, the first schedule allowing a first number of write operations over a period of time, the second schedule allowing a second number of read operations over the period of time, the second number being twice or more the first number; and a first port module is operable to read a first portion of a packet from one or more memory units for communication to a first component of the communications network before a second port module has received a second portion of the packet communicated from a second component of the communications network.

23. Logic for memory interleaving in a high-speed switching environment, the logic embodied in media and when executed operable to: couple a plurality of memory units to a plurality of port modules, each memory unit comprising one or more memory devices, each port module being operable to receive a packet communicated from a component of a communications network and write the received packet to one or more of the plurality of memory units, each memory unit being further operable to read a packet from one or more of the plurality of memory units for communication to the component of the communications network, the plurality of memory units being coupled to the plurality of port modules such that: each of the port modules is operable to write to each of the memory units according to a first schedule and each of the port modules is operable to read from each of the memory units according to a second schedule, the first schedule allowing a first number of write operations over a period of time, the second schedule allowing a second number of read operations over the period of time, the second number being twice or more the first number; and a first port module is operable to read a first portion of a packet from one or more memory units for communication to a first component of the communications network before a second port module has received a second portion of the packet communicated from a second component of the communications network.