Patent Grant
7054950
This description relates to computer networks.
Users of computer networks often want a rich variety of network services while also demanding high-speed data access. But network services take processing time that tends to slow the data delivery rate.
One way to achieve both high-speed and high-performance data packet processing is through parallelism or multi-threading in which network devices execute more than one thread simultaneously. Multi-threading is useful when a single task takes so long to complete that processing packets serially would slow down the overall packet throughput of a system too much.
During multi-threading, data packets sometimes flow into a multi-threaded task in one order (for example, P1, then P2) and flow out of the multi-threaded task in a different order (P2, then P1). This reordering can occur when the task is performed on P1 in one thread more slowly than the performance of the task on packet P2 in another thread. The different speeds of processing can result from the contention between threads for shared resources, such as memory bus access. Packet reordering is undesirable if the two data packets, P1 and P2, are bound for the same host.
Techniques that have been proposed to prevent packet reordering include adding hardware support, using only single-threaded processing, guaranteeing execution time per thread or using barrier or other synchronization mechanisms to ensure sequential processing of packets.
A basic mechanism for avoiding per-host packet reordering in a multi-threaded system ensures that packets en route to or from the same host will not be reordered even in the presence of semi-random (or even random) per-thread and per-task completion times.
As shown in
Referring again to
The characteristic identified by the first function 40 should distinguish which packets within the overall stream of packets 30 should not be reordered with respect to one another. For example, the first function 40 may characterize each data packet according to its destination address, thereby ensuring serial processing of all data packets bound for the same destination. Similarly, the first function 40 may characterize each data packet according to its source address, which, again, ensures that data packets delivered to an end host from any one source will remain in order.
In the example implementation of
Task A performs a four-bit masking of the least significant bits of the destination Internet Protocol (IP) address in each data packet. The four-bit mask produces a data packet identifier, H. Note that in this example, the data packet identifier, H, is an integer valued between 0 and 15, that effectively organizes the packets into 16 groups according to their destination addresses. The groups can then be distributed to the threads of any of the tasks according to the identifier H. To utilize all threads that make up any of the tasks in the processing of the packets, the number of bits chosen for the mask should correspond to at least the maximum value of the number of threads used for any one task. In the example of
Task A also stores the packet identifier, H, as an addendum to each packet by adding a byte of data containing H between each data packet in the stream of data packets. Therefore, the data packets may flow into Task A as P1, P2, P3, P4, P5, and P6 and leaving Task A as P1, H1, P2, H2, P3, H3, P4, H4, P5, H5, and P6, H6, where H1 is the data packet identifier for packet P1, H2 is the data packet identifier to P2 and so on. While the example illustrated in
After completing Task A, the data packets continue to flow to other tasks within the network device such as Task B. When the packets reach a task that immediately precedes a multithreaded task, for example, Task C, the task routes the packets to thread H Modulo K of the next Task D, where K is the number of threads in Task D. Threads 0–5 of Task D execute the same task (Task D) on several data packets at the same time. The data packets then flow out of the threads in an order determined by how fast the respective threads complete the task, and continue to the next task, Task E. After completing the series of tasks, Tasks A–E, the data packets are delivered to the end-host, 81, 82, 83, 84, 85, at the destination Internet Protocol address contained in each data packet.
The following chart illustrates the operation of the thread routing mechanism depicted in
In the above illustration, packets P1 and P4 in the data stream are bound for the same IP address (i.e., 123.5.17.6). Task A, the four-bit masking of the IP address, produces the same data packet identifier, H, for both packet P1 and P4. By having the same packet identifier, H, Task C will send P1 and P4 to the same thread, Thread 0 in this case. Since P1 and P4 are sent to the same thread, they are processed in series, and, therefore, will not be reordered relative to each other. This process ensures that the end-host will receive its data packets in the correct order assuming that no other intervening processing causes reordering.
Using a mask of the destination IP address, as illustrated by the four-bit mask in
In the alternate embodiment shown in
In addition to the implementation described above. Other implementations are also within the scope of the following claims.
For example, many other functions besides a mask or hash of the destination IP address may be used to produce a packet identifier. A mask or hash of some bits at any appropriate protocol layer could be used to distinguish packets and direct them to a particular thread. In the case of a Multi-Protocol Label Switching (MPLS) network, the label at the top of the stack could be used. In the case of an Ethernet network, the Media Access Control (source or destination) address associated with each data packet could be used. In the case of an system which uses the Asynchronous Transfer Mode (ATM) protocol, the Virtual Path Identifier (VPI) or Virtual Channel Identifier (VCI) may be used. Again, a mask or hash of some bits at any appropriate protocol layer could be used to distinguish packets and direct them to a particular thread and the technique described should not be limited to only internet, MPLS, Ethernet, or ATM networks.
Other techniques besides the modulo function may be used to route packets with the same identifier to the same thread. For example, a series of binary AND or other logic gates may be used.
H may be handled in a variety of ways. For example, it may be stored as an addendum to the data packet by adding a byte of data containing H in between each data packet (as described in the example illustrated in
The data processing technique described may be implemented in hardware, firmware, software, or a combination of them in network devices capable of parallel processing such as Network Processors with multiple microengines, multiple Network Processors, or multiple Application Specific Integrated Circuits.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5872783 | Chin | Feb 1999 | A |
| 5956721 | Douceur et al. | Sep 1999 | A |
| 6356951 | Gentry, Jr. | Mar 2002 | B1 |
| 6798743 | Ma et al. | Sep 2004 | B1 |
| 20030093555 | Harding-Jones et al. | May 2003 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20030195916 A1 | Oct 2003 | US |
