Patent Grant
7545808
The present invention relates generally to data routing systems, and more particularly to methods and apparatus for efficiently routing packets through a network.
In packet switched communication systems, a router is a switching device which receives packets containing data or control information on one port, and based on destination information contained within the packet, routes the packet out another port to the destination (or an intermediary destination).
Conventional routers perform this switching function by evaluating header information contained within a first data block in the packet in order to determine the proper output port for a particular packet.
Efficient switching of packets through the router is of paramount concern. Referring now to
Data packets received at an input port 2 are stored at least temporarily, in input buffer 4 while destination information associated with each packet is decoded to determine the appropriate switching through the switching device 6. The size of input buffer 4 is based in part on the speed with which the destination information may be decoded. If the decoding process takes too long as compared to the rate at which packets are received, large sized memory elements may be required or packets may be dropped.
In addition, the size of input buffer may be influenced by a condition referred to as “blocking”. Packets may be forced to remain in the input buffer after the destination information is decoded if the switching device cannot make the connection. Blocking refers to a condition in which a connection cannot be made in the switch due to the unavailability of the desired output port (the port is busy, e.g., routing another packet from a different input port). In summary, the size of input buffer 4 is dependent on a number of factors including the line input rate, the speed of the look-up process, and the blocking characteristics for the switching device.
Unfortunately, conventional routers are inefficient in a number of respects. Each input port includes a dedicated input buffer and memory sharing between input ports is not provided for in the design. Each input buffer must be sized to meet the maximum throughput requirements for a given port. However, design trade-offs (cost) often necessitate smaller buffers for each port. With the smaller buffers, the possibility arises for packets to be dropped due to blocking conditions. While excess memory capacity typically exists in the router (due to the varied-usage of the input ports), no means for taking advantage of the excess is afforded.
To minimize the occurrence of dropping packets, designers developed non head-of-line blocking routers. Referring now to
However, these designs are even more inefficient in terms of memory capacity and cost. Again, each output port includes a dedicated output buffer and memory sharing between output ports is not provided for in the design. Each output buffer must be sized to meet the maximum throughput requirements for a given port (in order to maintain its non head-of-line blocking characteristics). Even more excess memory capacity typically exists in the router (due to the varied usage of the input ports and output ports), yet no means for taking advantage of the excess is afforded. Twice the amount and bandwidth of memory has to be used than required to support the amount of data being moved through these types of devices.
One aspect is directed to a network device for switching a variable length data unit from a source to a destination in a network. The network device includes an input port for receiving the variable length data unit; a divider for dividing the variable length data unit into uniform length data units for temporary storage in the network device; a distributed memory including a plurality of physically separated memory banks addressable using a single virtual address space; and an input switch for streaming the uniform length data units across the memory banks based on the virtual address space. The network device further includes an output switch for extracting the uniform length data units from the distributed memory by using addresses of the uniform length data units within the virtual address space, the output switch reassembling the uniform length data units to reconstruct the variable length data unit; and an output port for receiving the variable length data unit and transferring the variable length data unit to the destination.
Another aspect is directed to a method of routing a variable length data unit through a network device in a system transmitting variable length data units between a source and a destination over a network including the network device. The method includes receiving the variable length data unit and dividing the variable length data unit into data units of a uniform size. The method further includes storing the uniformly sized data units in a distributed memory, the distributed memory including two or more memory banks where consecutive uniformly sized data units from the variable length data unit are stored in consecutive banks of the distributed memory. The method further includes storing linking information in one bank of the memory for uniform data units of the variable length data unit that are stored throughout the distributed memory, the linking information being used for extracting the uniform data units in order for transmission from the network device to the destination.
Yet another aspect is directed to a network device including an input port for receiving a variable length data unit, a divider for dividing the variable length data unit into uniform length data units for temporary storage in the network device, and a distributed memory comprising a plurality of physically separated memory banks. The network device further includes an input switch for streaming the uniform length data units across the memory banks: an output switch for extracting the uniform length data units from the distributed memory, the output switch reassembling the uniform length data units to reconstruct the variable length data unit; and an output port for receiving the variable length data unit and transferring the variable length data unit.
a and 1b are block diagrams of conventional router devices.
a is a schematic block diagram of a data routing system according to one embodiment of the present invention.
b is a schematic block diagram of a router according to one embodiment of the present invention.
a is a schematic block diagram of an multi-function port according to one embodiment of the present invention.
b is a schematic block diagram indicating data transfers between components of the router of
c is a data structure for a cell transferred between a multi-function port and an input switch according to one embodiment of the present invention.
a is a schematic block diagram of an input switch according to one embodiment of the present invention.
b is a schematic block diagram of a router including the timing and ordering of transfers from the input port to memory according to one embodiment of the present invention.
a is a schematic block diagram of main memory according to one embodiment of the present invention.
b is a schematic block diagram of a global data buffer according to one embodiment of the present invention.
c is a schematic block diagram of a router including 3 memory banks according to one embodiment of the present invention.
d is a schematic block diagram of a virtual address block for use in addressing memory according to one embodiment of the present invention.
e is a schematic block diagram of an index table for use in mapping physical address space to virtual address space according to one embodiment of the present invention.
f shows the relationship between physical and virtual addresses according to one embodiment of the present invention.
g is an example of a mapping between physical space and virtual space according to one embodiment of the present invention.
h shows the relationship between physical and virtual address space over time according to one embodiment of the present invention.
a is a schematic block diagram of an output switch according to one embodiment of the present invention.
b is data structure for a cell output from a memory bank to output switch according to one embodiment of the present invention.
a is a schematic block diagram for an output section of a multi-function port according to one embodiment of the present invention.
b is a schematic block diagram for a queue system for storing notifications according to one embodiment of the present invention.
Referring to
Referring to
In operation, packets are received at a multi-function multiport 150, transferred to input switch 100 and stored temporarily in global data buffer 104. When the packet is received by switch 100, a key is read from the first data block in the packet and transferred to controller 106. The key contains destination information which is derived from the header field associated with the first block of data in a packet and other information (such as source ID, priority data and flow ID).
A route look-up engine 110 in controller 106 performs a trie based search based on the key information and returns a result which includes the output multiport associated with the destination. The result is coupled with other information (such as source ID, flow ID and packet length) for routing the packet through router 20 and provided as a notification from controller 106 to output switch 102. Output switch 102 transfers the notification to the identified multi-function multiport 150. Upon receiving the notification information, the multi-function multiport 150 initiates the transfer of the packet from global data buffer 104 through output switch 102 to the appropriate multi-function multiport 150.
Multi-Function Multiports
Referring to
Input section 270 includes a line input interface 300, a data handler 302 and an input switch interface 304. Output section 280 includes an output request processor 306, a line output interface 308, a storage device 310, stream output buffers 312 (one for each output stream), output formatter 314, an output switch interface 316 and head and tail queue buffer 318. In addition, the output section includes a portion of input switch interface 304. Specifically, input switch interface 304 includes read request queues 305, one for each memory bank. The use and operation of the read request queues, stream output buffers, and head and tail queue will be discussed in greater detail below in association with
Memory section 290 includes a memory bank 105 (which represents a portion of the global data buffer 104) and a notification queue body 319. The use an operation of the memory section will be discussed in greater detail below in association with
The multi-function multiport is used in conjunction with the input switch, output switch and controller as is shown in
Referring again to
The format for transfers between the multi-function multiport and the input switch is shown in
The type field 356 indicates the type of cell to be transferred from the multi-function multiport. At each cell slot (20 clock cycles in one embodiment), a multi-function multiport may transfer either a data cell, an indirect cell placeholder, or a delayed indirect cell placeholder. Data cells contain data associated with an incoming packet. An indirect cell placeholder is an empty cell, and is used in conjunction with indirect addressing for the storage of the cells in the global data buffer 104. Delayed indirect cell placeholders arise when a data stream that requires indirect addressing terminates at a time prior to the designated time for writing the last indirect addressing cell associated with the data stream to global data buffer 104. The generation and operation of indirect placeholders and delayed indirect placeholders will be discussed in greater detail below in conjunction with
Stream field 358 indicates the stream to which the cell data belongs. In one embodiment of the present invention, each multi-function multiport is capable of handling up to sixteen separate streams of data at a time, one on each of its respective 16 input ports.
Packet header field 360 contains header information associated with a given packet and includes start offset information, packet length and interface index information.
Multi-function multiport identifier 366 identifies the multi-function multiport which is sourcing the read request. Address 368 indicates the address in global data buffer 104 to be read.
Referring now to
In one embodiment, cells from a given stream may be written to memory in an order that is different from the arrival order. These out of order writes are performed to make efficient use of scarce bandwidth between the multi-function multiports and the input switch. When a packet comes in to the multi-function multiport, it is broken up into cells as the bytes arrive and the cells are placed in per-bank output queues on the way to the input switch. These queues are designed to share scarce interconnect bandwidth between the streams of a multi-functional multiport in the most efficient way possible, but they have the detrimental effect of reordering cells at the interface between the multi-function multiport and the input switch. Thus the cells from a given stream may arrive at the input switch out of order. The multi-function multiport marks the data cells of a stream with one of four codes: first cell (FC); intermediate data cell (DC); last cell (LC); or first cell which happens to be also a last cell (FLC).
Input Switch
Referring to
Transfers Through the Input Switch
Round robin data handler 500 receives cells from each multi-function multiport and transfers them to output processor 505 for output to an appropriate memory bank 105 in global data buffer 104. Round robin data handler 500 services the inputs (cells) received on input interfaces 501 in a round robin, time division multiplexed manner. That is, for a given cell slot, one cell from each multi-function multiport is received at the round robin data handler 500 and subsequently transferred to output processor 505 for transfer at the next cell slot to a memory bank 105 in global data buffer 104. At the next time cell slot, data handler 500 transfers the next cell received from the same multi-function multiport to output processor 505 for transfer to a different memory bank. In one embodiment, the next cell received is transferred to the next memory bank (next in numerical order, modulo N) in the memory array. Alternatively, another time dependent permutation may be used to control the transfer of successive cells from the same multi-function multiport.
Referring to
Round robin data handler 500 and output processor 505 within the input switch 100 transfer cells out to global data buffer 104 on transmission lines 460. As can be seen at cell slot T2, output processor 505 outputs one cell 454-B0 to 454-B7 to each memory bank in a single cell slot. The “Bx” designator indicates the multi-function multiport from which the particular cell was received. One cell from each multifunction multiport is written to global data buffer 104 per cell slot. At time period T1 (one cell slot prior to cell slot T2), again one cell (456-B0 to 456-B7) is written to each memory bank. Round robin data handler 500 time division multiplexes the transfers to output processor 505 such that consecutive cells from the same multi-function multiport are written to consecutive memory banks 105 (modulo N) in global data buffer 104.
Referring again to
Key Reading and the Linking Process
Round robin data handler 500 includes a key reading engine 514 for determining the key information associated with a first cell in a packet and a linking engine 515 for linking cells in the same packet.
The process of reading key information is known in the art. After the key is determined for a given packet, it is stored temporarily in key buffer 516 in input switch 100 until the entire packet has been stored in global data buffer 104. The data structure for entries 600 in the key buffer 516 is shown in
Referring now to
If more than five data cells are included in a packet, then the indirect cell indicator for that packet is set, and the last offset indicates the address in memory where the first indirect cell associated with the packet is stored. Indirect cells will be described in greater detail below in reference to
The linking or threading of cells for a packet is performed by using the offsets described above and indirect cells. Offsets are used to link cells in a packet. Offsets may be stored along with key information and routed through controller 106 (
Referring now to
As was described above, when a packet is received, the linking engine processes the first five cells and stores linking information in the form of a start address and four offsets in key buffer 516. In the event more than five cells are contained within a packet, the indirect cell processor takes over for the linking engine and computes the offsets associated with the locations in memory where the remaining cells in the packet are stored. Round robin processor 500 passes cells to the output processor 505 for transfer to an associated memory bank in global data buffer 104. Round robin processor 500 enables the indirect cell processor when the packet being processed contains more than 5 cells (based on header information included within the first cell). At the time for writing the fifth cell to memory, indirect cell processor 510 stores in indirect cell memory 520 the address (the “indirect cell address”) associated with the location in memory at which the fifth cell would have been written if it had been the last cell in the packet. The indirect cell address indicates the location in memory where the indirect cell is to be written-when full (or when the last cell of the packet is processed).
When an indirect cell is full (having stored offsets in all available locations except the last field 706), then the indirect cell processor stores the offset associated with the location in memory where the next indirect cell is located in the last field 706. Thereafter, the full indirect cell is written to its appropriate-place in memory. The writing of the indirect cell to memory coincides with the receipt of an indirect cell placeholder by the input switch 100 from the associated multi-function multiport 150. This process continues until the last cell in a packet is stored in memory. At that time, the last indirect cell is written to memory, and the associated entry 600 from the key buffer 516 is transferred to the controller 106 for processing. For a given packet, all indirect cells are written to the same memory bank in the global memory buffer.
As often will be the case, the last cell of a packet will not coincide with the timing required to write the completed indirect cell immediately into memory. This is because packet length is completely arbitrary. The end of a packet will likely not coincide with the last-available entry of an indirect cell. When a packet has completed (all cells have been received by the input switch) and a last entry in the indirect cell is written, the indirect cell is free to be written to memory. However, the writing will be delayed until the proper time, hence the term delayed indirect cell. A delayed indirect cell is a indirect cell that is the last indirect cell associated with a packet. It is delayed, because it is written to memory after the rest of the packet has been written to memory. The timing of the write to memory is dictated by the address which is reserved for the indirect cell. As was described above, at the time for the creation of an indirect cell, its position in memory is reserved. The delayed indirect cell will be written to memory at the next time slot available for the particular multi-function multiport to write to the particular memory bank after the packet has been completed. The timing of the write to memory of delayed indirect cells coincides with the receipt of a delayed indirect placeholder from the appropriate multi-function multiport 150.
Transfers to Memory
The data structure of a cell transferred from input switch 100 (via the output processor 505) to a memory bank 105 in global data buffer 104 is shown in
At each cell slot, output processor 505 generates a cell 800 which includes a read request source field 802, read address 804, write address 806 and data field (cell data received from multiport 150) 808. The read request source field 802 indicates the output port (in the particular multi-function multiport 150) requesting the read (destination output port). Output processor 505 receives read requests from read controller 517 and bundles the read request with any write request received from round robin data handler 500 destined for the same memory bank. At each cell slot, output processor 505 provides a cell 800 which may include a write and read request to each memory bank 105 in global data buffer 104.
Read controller 517 controls the transfer of read request signals flowing from input switch 100 out memory interface 502 to the individual memory banks in global data buffer 104. Read controller 517 receives read requests from each multi-function multiport through output interfaces 506. The format of each request includes source identification (output port) and a full address in memory which is to be read. At each cell slot, each multifunction multiport port may generate a read request for processing by switch 100 to read a memory location in global data buffer 104, resulting in the reading of a cell (a read reply) from a memory bank 105 (on a subsequent cell slot) to output switch 102.
Read controller 517 loads a reservation table 508 as requests to transfer packets are received from the various multi-function multiports 150. The reservation table is loaded such that at every cell slot a single read request is generated for each bank of memory 105. Referring now to
Loaded entries 906 reflect read requests to be performed as a result of reservation requests received from individual multi-function multiports. Placeholders 904 represent available slots in which read requests for a particular memory bank are still available (e.g., read requests which have not as of yet been received for this memory bank from a particular multi-function multiport). At each cell slot, the read controller 517 performs three functions: loading entries in the reservation table at the first available location in the table (after the read pointer), outputting the last row as read requests to the output processor 505; and refreshing the table, moving out the last row, incrementing the rows and creating a new row at the top of the table. The number of rows in the reservation table must be as large as the product of the latency in processing read requests multiplied by the number of banks. In one embodiment, 48 rows are included in reservation table 508 reflecting a system including six cell-slots of latency and eight memory banks.
At initialization, reservation table 508 contains placeholders 904 in all of the rows 902. Placeholders 904 are locations in the reservation table which have not been loaded. As read requests are processed by the read processor, certain ones of the placeholders 904 are converted to loaded entries 906 based on the read requests. Loaded entries 906 include a read request address.
Referring now to
Memory Architecture
Referring now to
Global data buffer 104 includes a plurality of memory banks 105. Associated with each memory bank is an input switch interface (an input port) 304 and output switch interface (an output port) 316. At each cell slot, each memory bank receives at most one write and one read request via input switch interface 304. The write requests are associated with cells received from a multi-function multiport 150. Read requests reflect a request for cell data to be transferred from a memory bank 105 to output switch 102 for ultimate transfer to a requesting multi-function multiport 150.
The memory in the multi-function multiport configuration is physically distributed across a number of banks b, one bank for each active multi-function multiport in the system.
Each bank=s memory bandwidth is sufficient for reading and writing packets from a full-duplex OC-48 interface as well as for queuing and dequeuing notifications for the worst-case example of single-cell packets. Thus, both the aggregate memory size and the aggregate memory bandwidth scale linearly with the number of active multi-function multiports b in the system.
In one embodiment, each memory bank is implemented as two sub-banks using two 72-bit wide SDRAM (static dynamic random access memory) DIMM=s (dynamic in-line memory modules) cycling at 125 MHZ. The sub-banks are transparent to the input and output switch resulting in what appears to be one continuous bank from the perspective of the switches. However, the sub-bank architecture allows for better throughput. Each DIMM has a 72-bit wide ECC (error correction code) protected data path going to 9 SDRAM chips each of which is 8 bits wide. The two DIMM=s have separate address busses and are addressed independently of one another. The DIMM=s are interleaved on bit 0 of the 23-bit address. In one embodiment, the smallest memory bank configuration is 32 MBytes, using 16 Mbit chips and the largest is 512MBytes, using 256 Mbit chips.
As was described above, a bank can receive at most one read request and one write request every cell slot. Since a cell slot is 20 clock cycles at 125 MHZ, this works out to a peak bandwidth demand of 400 MBytes/sec for reads and 400 MBytes/sec for writes. The worst case notification load occurs for single cell packets. For unicast traffic, this load is exactly ¼ the data bandwidth which works out to 100 MBytes/sec for reads and 100 MBytes/sec for writes. In this embodiment, the total peak memory bandwidth needed is therefore 1 GByte/sec.
In this embodiment, the peak transfer rate of each DIMM is 1 GByte/sec, but the sustained rate depends on the actual mix of reads and writes and how the addresses are distributed over the internal DIMM banks. In practice, each DIMM is expected to deliver a sustained data rate of around 650 MBytes/sec. The total of 1.3 GBytes/sec supplied by the two groups is 30% larger than the maximum sustained requirement of 1 GByte/sec/. The 30% headroom provides a way to sustain instantaneous loads where one DIMM has more references directed to it than the other. The memory controller for the two DIMM=s resides in the multi-function multiport.
In one embodiment, all banks are made the same size and approximately ⅕th of the memory in each bank is allocated to the notification area and ⅘th to the global data area. The purpose of this allocation is to make it exceedingly unlikely for a stream to run out of memory because of space in its notification queue. With a worst case packet size of 64 bytes, notifications (sized at 16 bytes) need ¼th the amount of storage that packet data needs, which is exactly the proportion allocated. Any cell in the global data buffer may be accessed via its physical cell pointer, which identifies the physical bank number and the address of the cell within the bank. The physical cell pointer defines a system-wide physical address space. To simplify address computations, as well as to provide a mechanism to detect old packets, accesses to the global packet buffer are performed through a system-wide virtual address space that maps to the physical address space.
Incoming packets are broken up into as many cells as needed and the cells are written to the global packet data buffer as they arrive as described above. The global data buffer is treated as a single large circular buffer. The input switch maintains an array of write pointers, one per active bank, to keep track of where to write the next cell. The pointers start out at one end of the buffer and advance until they eventually wrap around and overwrite packets that were written a long time ago. An ageing mechanism is used to guard against reading cells that may have been overwritten by subsequent packets. The cells of packets arriving on a given stream are interleaved strictly across the active banks to spread the bandwidth load.
Addressing
Global data buffer 104 is divided into banks which are in turn divided into fixed size, non-overlapping 64-byte units called cells. The cell is both the unit of addressing and of memory allocation. All accesses to memory are performed using two operations: cell read and cell write. Since there is only one access size, memory allocation is straightforward and there are no issues of data alignment or byte order.
Each cell in the system is uniquely identified by a physical cell pointer or PCP, which identifies the cell's physical bank number and address within that bank. The set of PCP 's form a system-wide physical address space that can be used to read or write any cell in the system. While this space provides access to all possible banks in main memory, it is not convenient for reading and writing cells that are part of the global data buffer.
For example, physical space may have holes in it corresponding to memory banks (multi-function multiports) that are either not plugged in or are plugged in but inactive.
In addition, the size of the global data area is not necessarily a power of two (since an odd number of multi-function multiports may be available and because of the ⅘ to ⅕ ratios set forth for the global area and notification areas respectively), so address computations involve checks that would normally not be needed. Both these factors make the address arithmetic that needs to be performed on each access awkward, and using physical space directly forces each entity issuing reads and writes to know how to do this arithmetic. Finally, direct access to physical space provides no mechanism for detecting if a notification is pointing to cells that have been overwritten since the original data was written to memory. A virtual-to-real address mapping mechanism provides a way to solve these problems. This mechanism defines a single system-wide virtual address space that allows the global packet buffer to be read and written conveniently. All non-diagnostic accesses to the buffer use this address space.
Physical Space
Each memory bank has a unique 3-bit physical bank number, or PBN, that is equal to the number of the slot in which the bank is plugged. In one embodiment, slots are numbered sequentially from 0 to 7 (assuming an eight multi-function multiport configuration). Note that a bank may be plugged into any available slot so that the set of active bank numbers in a system can be an arbitrary subset of the numbers [0 . . 7]. The PBN serves as the physical address of a bank.
Within a bank, cells are identified using a 23-bit physical cell address, or PCA (1124 of
A bank's memory is divided into two contiguous non-overlapping regions: cells [O. . Nbuf−1] are allocated to the global packet buffer, while cells [Nbuf. . Nbank] are allocated to queues that hold notifications for packets to be sent out via the associated multi-function multiport. The number of cells Nbuf allocated to the global data buffer is the same for each of the b active banks so each bank contributes exactly 1/bth of the storage for the buffer. Typically, around ⅕ of the memory in a bank is allocated for notification queues and ⅘ for the global data buffer.
Although each bank is shown to be the same size, note that only the global packet buffer areas are required to be the same size across banks. The architecture places no such restriction on the size of the notification areas even-though they will typically be the same size.
Any cell in the system is identified uniquely by the 26-bit tuple <PCA><PBN>. This tuple is called the cell's physical cell pointer, or PCP. The set of possible PCP values defines a system-wide physical address space. All cell references must be translated to a PCP before the cell can be read or written. In one embodiment, the maximum amount of memory in the system is limited by the architecture to 226 cells, or 4 GBytes.
Virtual Space
Referring now to
The set of values for VBN define a virtual bank number space that make it convenient to do arithmetic on bank numbers. If there are b active banks in the system, the legal values of VBN are [0 . . b−1], and all arithmetic on VBN's is done mod b.
The set of values for VCA define a virtual cell address space for the bank identified by VBN. This virtual cell address space makes it convenient to do arithmetic on cell addresses that lie in its bank and also enable ageing checks.
A VCP is translated to a physical cell pointer (PCP), by performing two separate mappings: first, the VBN is mapped to a physical bank number (PBN) and then the VCA is mapped to a physical cell address (PCA). The two mappings are not independent. In particular, the VCA to PCA mapping uses states particular to the physical bank produced by the first mapping.
Bank Number Mapping
Virtual bank numbers are translated to physical bank numbers by the input switch using an 8-entry lookup table 1150 as shown in
In general, if there are b active banks in the system, the first b entries of the table are filled with the physical bank numbers for these banks.
Cell Address Mapping
Referring now to
For purposes of simplifying mapping, the virtual and physical cell address spaces are broken up into contiguous, self-aligned, 4096 cell regions called pages. The low-order 12 bits of both VPA's and VCA's correspond to the address of a cell within a page, while the high-order bits correspond to the address of the page itself. Referring now to
The operation of the VPMap (1130) function is as follows: Let N be the number of pages contributed to the global packet data buffer by each bank. From the structure of VCA's, it is clear that this number must be an integer in the range [1 . . . 2048] (The value 0 is excluded because it represents the case of no buffer memory). Note that the number of pages in virtual space is exactly two times the largest value of N. As will be seen later, this fact is crucial for the ageing scheme to work.
The VPMap function is time-dependent, but its operation at a given time is straightforward. For each bank this function maps a contiguous region of N pages in virtual space [νTP, νTP+1, . . . , νLP] onto the N pages of physical space for that bank (where TP stands for “trailing page” and LP for “leading page”). All regions are located at the same point in their respective spaces at a given time. For this discussion, all arithmetic is mod 4096, so virtual address space is assumed to wrap around after page 4095. One of the virtual pages in the region is always mapped to physical page 0. Let this page be ν0=νTP+k. Then, virtual pages [vTP+k, vTP+k+1, . . . , νLP] are mapped directly to physical pages [0, 1, . . . , N−k−1], and virtual pages [νTP, vTP+1, . . . , νLP+k−1] are mapped directly to physical pages [N−k−1, N−k, . . . , N−1]. That is, the N page region is broken into two disjoint pieces and pages in the two pieces are mapped one-for-one to identically sized pieces of physical space. References to virtual pages that lie outside the region are mapped to physical page 0xFFF to indicate that the reference was to a cell that has been overwritten.
The time-dependent behavior of VPMAP is that the N page region of virtual space that is mapped moves with time. At system initialization VLP is set to 0, and VTP is set to −(N−1). As cells are written to system memory, the write pointers for each bank advance more or less together but they do not move in lock-step. The system computes VLP as the high-order page bits of the bank write pointer that is furthest ahead; VTP then is simply (VLP−N+1).
As the region advances, the value of v0 must be updated periodically so it stays in the range of virtual addresses that is being mapped. At system initialization v0 is set to 0, and subsequently each time v0 becomes equal to vTP, it is incremented by N. The increment is done after mapping, not before.
Transfers from the Input Switch to Memory
Referring now to
Controller
Referring now to
The route look-up engine servicing the route look-up request performs a best match look-up and outputs a notification through output switch interface 1302 to output switch 102. The notification includes a result which indicates the multi-function multiport to be used in the transfer of the packet to its destination.
Referring now to
The mask field 1402 is used to indicate which multi-function multiport connected to output switch 102 is to transfer the packet. In one embodiment, the notification may be sent to more than one multi-function multiport resulting in the broadcast of the associated packet.
Associated with each multi-function multiport 150 is a storage 310. The next hop index pointer points to a location in storage 310. Storage 310 is used to store media header information associated with a particular type of packet transfer. Next hop addresses, media headers and storage 310 will be described in greater detail below in association with the output section of multi-function multiport 150.
The full address 1406 indicates the starting address in the global data buffer where the first cell in the packet is stored. As was described above, offsets 1408 provide linking information for retrieving cells or an indirect cell associated with the packet. Packet length filed 1410 indicates the length of the associated packet and may be used to determine if indirect cells will have to be retrieved.
Output Switch
Referring now to
Transfers from Memory to the Output Switch
Cells from memory are received at memory inputs 1502 and transferred to output processor 1508. Cells are transferred based on read requests received at the input switch from multi-function multiports. The data structure associated with the cells transferred from global data buffer 104 to output switch 102 is shown in
Output processor 1508 decodes the destination multi-function multiport from the cell information received from memory and transfers the cell data to the appropriate outputs 1502. At each cell slot, output switch 102 may receive a cell for processing from each bank in global data buffer 104.
Transfers from the Output Switch to the Multi-Function Multiports
Output switch 102 receives notification from controller 106 on controller interface 1500. Result processor 1506 decodes the result (route) and determines which multi-function multiport(s) 150 is (are) to receive the route data. Based on mask 1402 in the notification, result processor 1506 transfers the notification to output processor 1508 for transfer to each multi-function multiport 150 indicated. At each cell slot, output processor 1508 provides (via outputs 1504) a route to each multi-function multiport 150.
The data structure associated with the data transferred from output processor 1508 to multi-function multiports 150 is shown in
Output Section of a Multi-Function Multiport
Referring now to
Notification Queues
A multi-function multiport 150 receives notification that a packet is to be processed in the form of a notification cell 1600 (
Output request processor 306 processes notifications, storing each in an appropriate location in head and tail queue buffer 318 and servicing notification requests as they make their way through the various priority queues in head and tail queue buffer 318. The servicing of requests results in the generation of a read request to input switch 100 associated with the first address in memory where the packet (associated with the particular notification) is stored.
Referring now to
The size of the portion of the head and tail queue buffer dedicated to each stream is fixed at initialization time and is proportional to the peak bandwidth of its stream. The partitions between portions are “hard” in the sense that a stream cannot use more than the memory allocated to it. The partitions between queues associated with the same stream are “soft”. The size of an individual queue is proportional to the nominal bandwidth allocated to its queue. The body region of the notification queue is stored in the notification area 319 (
The body region is sized to be ⅕ of the overall memory section.
Each queue associated with a given stream is serviced according to a priority scheme. Notifications that are received by the output request processor 306 are loaded into an appropriate queue associated with a stream based on the priority of the notification. Priority for notifications can be set by an external source and may be included in the packet received by the router. Alternatively, controller 106 (
Once a queue 1700 has been identified based on the priority information and stream ID, the output request processor 306 loads the notification into the appropriate tail queue 1704. Notifications are transferred between the respective head, tail and body portions of a queue based on available space by a queue manager (not shown). In one embodiment, each notification is 16bytes, and the notification area 319 is sized to hold 64 bytes. Accordingly, for reasons of bandwidth efficiency, all reads and writes to the notification area are done using 64-byte cells containing four 16-byte notifications each. The head and tail of each queue is sized to store only a small number of notifications, the bulk of queue storage being provided by the notification area in the multi-function multiport memory bank. As long as space is available on-chip (on the multiport) to hold the notifications for a queue, the notification area is completely bypassed. When on-chip space runs out, the notification area acts as the large “middle” of the queue, with a few notifications at the head and tail being held on-chip.
While the size of the notification area will tend to limit the numbers of dropped packets, occasionally a queue will become full. Output request processor includes a drop engine (not shown) for determining which entries in a particular queue are to be dropped based on a predefined algorithm. In one embodiment, the drop engine institutes a programmable random early drop routine. The routine is programmable in that the user can define one or more parameters, random in that a random number generator is used to determine whether a entry will be dropped. Early refers dropping from the head of the queue.
The programmable random early drop routine may be implemented in software and when executed performs the following sequence of operations. The process begins by calculating the amount of data stored in a particular queue. This information is stored in the form of a fraction (or percentage) of fullness. Thereafter, a drop criterion is determined based on the fraction of fullness. In one embodiment, a table of drop criterion values ranging from zero to one is mapped against fullness fractional values. The drop engine then derives a random number from zero to one. The random number may be generated by a random number generator or other means as is known in the art. A comparison is made between the random number generated and the drop criterion value. Thereafter, the entry at the head of the particular queue is dropped if the random number generated is larger than the drop criterion. Alternatively, the drop engine could avoid the drop if the random number generated is less than the drop criterion. The drop engine operates on each queue at a regular interval to assure that the queues do not overflow and a orderly method of dropping packets is achieved if required. This process is extremely helpful when transmitting packets across the Internet.
Per Bank Notification Queues
Each stream includes four queues 1700 that are serviced using a weighted round robin discipline. The weighting is used to reflect the priority associated with a given queue. For example, the four queues for a given stream may be serviced in the following ratios: Q1 at 50%, Q2 at 25%, Q3 at 15% and Q4 at 10%.
The multi-function multiport maintains four cell pointers for each queue: start, end, head, and tail. In one embodiment, each pointer is 23 bits long and can address the entire memory associated with the multi-function multiport. The start and end pointers mark the boundaries of the queue's region, while the head and tail pointers point to the next cell (notification) to read and next cell o to write respectively. The head and tail pointers are restricted to align within the region defined by the start and end pointers, and standard wraparound arithmetic is performed when incrementing these pointers.
Given the description above, it should be clear that the region for a queue can be as small as one cell and as large as the entire memory bank. It is up to the software to configure the pointers at initialization time to define the sizes of the regions; and to ensure that regions are non-overlapping with each other and with the memory allocated to the global packet buffer.
Typically, the software is used to allocate memory to a stream proportional to the stream's bandwidth.
Read Request Generation
Output request processor 306 services the queues to extract notifications from the head regions of queues 1700. Output request processor generates a first read request based on the full address received from output switch 102. Thereafter subsequent read requests are generated for transmission to the input switch based on the offset information provided in the request (from cell 1600) or indirect cells (as will be described below). Read requests include a stream identifier and a full address. Read requests are sent by the output request processor to an appropriate read request queue 305. One read request queue 305 is provided for each bank of memory.
If the packet length, as determined from the route information provided with the cell 1600, is greater than five (5) cells, then the output request processor first requests the transfer (read from memory) of the first indirect cell associated with the packet. This is accomplished by computing the address of the indirect cell based on the full address and the offsets provided in cell 1600. After the indirect cell request is generated, the output request processor generates read requests for the remaining cells in the packet based on the full address and the offsets provided in cell 1600. Upon receipt of a indirect cell from the output switch 102, output request processor continues to generate read requests for the remaining cells in the packet based on the offset information contained within the indirect cell.
Subsequent indirect cells are retrieved in a similar fashion. That is, at the time for reading the next indirect cell, the address of the next indirect cell is computed based on the last offset stored in the previous indirect cell. The timing of retrieving the indirect cells is accomplished such that no delays in the output stream are incurred. Each subsequent indirect cell is retrieved prior to the end of the processing of the prior indirect cell. In this way, once the output stream is initialized, no buffering of data is required and no interruptions due to the latency associated with the retrieval process are experienced.
Output requests to an individual memory bank are processed strictly in order. That is, the multi-function multiport may track each request issued to a memory bank (through the read request queues) and is assured that the data received in response to a series of requests to the same memory bank will be strictly delivered according to the sequence or pattern in which they were issued. Output request processor 306 keeps track of requests generated for each memory bank through the use of reply queues (not shown). The request queue contains a stream number and a read address. When a request is issued to memory, the entry is removed from the request queue and the stream number portion is placed in an associated reply queue. When a reply is received, the entry at the head of the reply queue is removed and the reply is sent to the stream number (in stream output buffer 312) indicated by the stream number retrieved from the reply queue.
As cells are received back at the multi-function Multiport 150 (responsive to the read requests), they are stored in an associated stream output buffer 312. Stream output buffer 312 includes a plurality of FIFOs, one for each stream. Each cell received for a stream is placed in the streams associated FIFO. For given packet, the multi-function multiport stores a fixed number of cells (in the FIFO) required to provide a streamed output prior to initializing the output of the stream to line output interface 308. In one embodiment of the present invention, twelve cells are stored prior to beginning output (stream data) from the output port. The selection of the number of cells for storage in output buffer 312 is based on the latency in the read process (number of clock cycles between a read request from an multi-function multiport and the arrival of the cell associated with the read request to the output port).
Output formatter 314 receives the cells from output buffer 312 and couples the data with media header information stored in memory 310. Each request (notification) received from output switch 102 includes a next hop index. The next hop index indicates the starting address in memory 310 of the media header information associated with a given type of transmission (derived from the destination of the packet). Media header information stored in memory 310 may be loaded upon initialization of the router and updated by the controller as required. Output formatter 314 couples the cell data returned from memory with the appropriate media header to generate a proper packet for transfer out of router 20 on the line output interface 308.
Packet Routing Overview
Referring now to
When the packet (and its indirect cells if any) have been stored in memory, then the key, full address of the first cell and offset information is transferred as a look-up request to a controller (1814). The controller performs a best match look-up and generates a result of the look-up. The result includes the destination port (multi-function multiport), address, offset information and next hop index (1816). A notification including the result is transferred to the output switch for transfer to the appropriate multi-function multiport (1818).
Upon receipt of a notification, the multi-function multiport generates read requests a cell at a time to the input switch for the data associated with the packet (1820). The input switch issues the read requests in a time division multiplexed fashion generating a single request to each memory bank per cell slot (1822). When the memory bank receives the request from the input switch, cell data and a multi-function multiport identifier associated with the request are transferred to the output switch (1824). Again, at each cell slot, the output switch transfers a single cell to each of the multi-function multiports. Upon receipt, the multi-function multiport couples the cell data with media header information and streams the data to the destination (1826).
System-Level View
Up to S streams, where S=s*b+1, may be in process in the system concurrently for b active banks (+1 is for the stream corresponding to the control channel (the control channel stream is used to control configuration and operation of elements in the router). Each stream consists of a sequence of packets of arbitrary length L with arbitrary inter-packet gaps. In one embodiment L is less than 64 KBytes. A stream operates at a constant rate R Bytes/sec when it is receiving a packet, and a rate of 0 Bytes/sec when it is not. Thus, there are no “holes” during the reception of a packet. The value R for a stream is determined at system initialization time and Rmin and Rmax are system constants defined at design time. In one embodiment, Rmin is approximately 4 MBytes/sec and Rmax is approximately 320 MBytes/sec.
As described earlier, each packet is divided up into 64-byte cells before being written to memory. If all packets are long compared to 64 bytes, then the rate of cells generated by a stream is simply R/64 or 0 depending on whether a packet is being received or not. Unfortunately, because real life packets do not satisfy this assumption, the instantaneous rate of cell generation depends on the lengths of the packets being received, and this rate will typically be larger than R/64 because of unoccupied bytes in cells. Because of this dependency, the most we can say is that a stream is either generating cells at some rate between R/64 and a small factor ƒ times this rate, or it is not. The worst-case value of ƒ for IP is 128/65, which occurs for 65 Byte packets where one cell is full and the other carries a single byte. For realistic network traffic, however, the average value of ƒ should be around 1.2, with bursts of 40 Byte packets causing ƒ to go as high as 1.6 for short periods.
Let the cells generated by a given stream be numbered I, I+1, I+2, . . . etc. As was described above, cells are written to sequentially increasing bank numbers I mod b. The virtual address to which a particular cell is written is determined largely by the cell's arrival time into the system. For each active memory bank, the system maintains a write pointer that determines where the next cell directed to that bank will be written. The pointers are kept in the input switch, and so is the logic to switch incoming cells to the appropriate bank. When the input switch receives a cell, it knows which bank the cell should be written to by the cell's ordinal number. The input switch looks up the write pointer for the destination bank, writes the cell, and then increments the bank pointer to point to the next location.
At system initialization, all bank pointers in the input switch are set to 0, and the S streams in the system are initialized with starting bank numbers such that approximately the same number of streams are pointed to each of the b active banks. Although the starting bank numbers are distributed evenly, they do not have to be. Distributing them evenly simply makes the magnitude of the maximum negative offset a little smaller.
Over time, the bank pointers move as cells are written to the buffer, but they move in a compact group because each stream writes its packets to sequentially increasing bank numbers mod b. The overall speed of the group is determined by the sums of the instantaneous bandwidths on all the streams, and can vary from 0 cells/sec when no packets are being received to a maximum of 400/64=6.25M cells/sec when all interfaces are running at peak capacity.
In an alternative embodiment, multi-function multiport 150 may be divided into separate functional blocks and embodied in separate physical structures. For example, each multi-function multiport may be separated into an input port, an output port and a memory. Memory may include both the global data buffer and the notification queues. Alternatively, memory may be divided further with the notification queues physically manifested as a part of an output port.
The present invention has been described in terms of specific embodiments, which are illustrative of the invention and not to be construed as limiting. Other embodiments are within the scope of the following claims.
This application is a continuation of U.S. patent application Ser. No. 10/310,462 filed Dec. 4, 2002, now U.S. Pat. No. 7,116,660 which is a continuation of application Ser. No. 08/901,061, filed Jul. 24, 1997 (issued as U.S. Pat. No. 6,493,347), which is a continuation-in-part of U.S. application number 08/844,171, entitled “HIGH SPEED SWITCHING DEVICE”, filed Apr. 18, 1997 (issued as U.S. Pat. No. 5,905,725), which is a continuation-in-part of U.S. application Ser. No. 08/767,576, entitled “HIGH SPEED VARIABLE LENGTH BEST MATCH LOOK-UP IN A SWITCHING DEVICE”, filed on Dec. 16, 1996 (issued as U.S. Pat. No. 5,909,440).
| Number | Name | Date | Kind |
|---|---|---|---|
| 4058672 | Crager et al. | Nov 1977 | A |
| 4596010 | Beckner et al. | Jun 1986 | A |
| 4884264 | Servel et al. | Nov 1989 | A |
| 4885744 | Lespagnol et al. | Dec 1989 | A |
| 4926416 | Weik | May 1990 | A |
| 4933932 | Quinquis et al. | Jun 1990 | A |
| 4947388 | Kuwahara et al. | Aug 1990 | A |
| 5155834 | Ryan et al. | Oct 1992 | A |
| 5214639 | Herion | May 1993 | A |
| 5293597 | Jensen et al. | Mar 1994 | A |
| 5448702 | Garcia, Jr. et al. | Sep 1995 | A |
| 5457687 | Newman | Oct 1995 | A |
| 5491694 | Oliver et al. | Feb 1996 | A |
| 5521910 | Matthews | May 1996 | A |
| 5521915 | Dieudonne et al. | May 1996 | A |
| 5668968 | Wu | Sep 1997 | A |
| 5689505 | Chiussi et al. | Nov 1997 | A |
| 5796944 | Hill et al. | Aug 1998 | A |
| 5802052 | Venkataraman | Sep 1998 | A |
| 5872787 | Cooperman et al. | Feb 1999 | A |
| 5905725 | Sindhu et al. | May 1999 | A |
| 5909440 | Ferguson et al. | Jun 1999 | A |
| 5956751 | Lai | Sep 1999 | A |
| 6201813 | Klausmeier et al. | Mar 2001 | B1 |
| 6493347 | Sindhu et al. | Dec 2002 | B2 |
| 7116660 | Sindhu et al. | Oct 2006 | B2 |
| Number | Date | Country | |
|---|---|---|---|
| 20060023719 A1 | Feb 2006 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10310462 | Dec 2002 | US |
| Child | 11226335 | US | |
| Parent | 08901061 | Jul 1997 | US |
| Child | 10310462 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 08844171 | Apr 1997 | US |
| Child | 08901061 | US | |
| Parent | 08767576 | Dec 1996 | US |
| Child | 08844171 | US |
