1. CSC457 Seminar YongKang Zhu December 6 th, 2001 About Network Processor

2. Outline 1. What is a NP, why we need it and its features 2. Benchmarks for NP evaluation 3. Several issues on NP design (a). Processing unit architecture (b). Handling I/O events (c). Memory (buffer) organization and management

3. What is a network processor? A network processor is a highly programmable processor, which is suitable for performing intelligent and flexible packet processing and traffic management functions at line speed in various networking devices, such as routers and switches, etc.

4. A typical router architecture

5. Why NP and their features?  Fast growth in transmission technology  Advanced packet processing functions  Traditional methods: using ASIC or off-the-shelf CPU  Performance  Programmability, flexibility  Design and implementation complexity  Value proposition

6. Benchmarks for NP evaluation Major metrics include: Throughput: bps, pps, connections per second, transactions per second Latency: time for a packet passing through NP Jitter: variation in latency Loss Rate: ratio of lost packets

7. Commbench - by Mark Franklin 1. Two categories of typical applications: Header processing applications: RTR, FRAG, DRR, TCP Payload processing applications: CAST, REED, ZIP, JPEG 2. Selecting appropriate input mix to represent different workload and traffic pattern 3. Design implications ( computational complexity )

8. Importance of selecting input mix

9. Some Issues on NP design  Processing unit architecture  Fast handling I/O events  Memory organization and management

10. Processing unit architecture Four architecture reviewed: 1. a super scalar microprocessor (SS) 2. a fine-grained multithreading microprocessor (FGMT) 3. a chip multiprocessor (CMP) 4. a simultaneous multiprocessor (SMP)

11. Comparison among four architectures 1. CMP and SMP can explore more instruction level parallelism and packet level parallelism 2. However, other problems are introduced, as how to efficiently handling cache coherency and memory consistency

12. Handling I/O  Make equal sized internal flits  Higher level pipeline for packet processing  Using coprocessor

13. Higher (task) level pipeline

14. Memory organization & management 1. Using novel DRAM architectures:  page mode DRAM  Synchronous DRAM  Direct Rambus DRAM 2. Using slow DRAM in parallel:  Ping-pong buffering  ECQF-MMA (earliest critical queue first)

15. Ping-pong buffering Buffer Organization Buffer Usage

16. ECQF-MMA (earliest critical queue first)  Using slow DRAM and fast SRAM to organize buffer structure  total Q FIFO queues  memory bus width is b cells  memory random access time is 2T  the size of each SRAM is bounded to Q * (b - 1) cells  Arbiter selects which cells from which FIFO queue will depart in future  requests to DRAM for replenishing SRAM FIFOs are sent after being accumulated to a certain amount  guarantee a maximum latency experienced by each cell

17. Intel's IXP1200  1 StrongArm core and 6 RISC micro engine  can manage up to 24 independent threads  two interfaces: IX bus and PCI  IX bus for connecting MAC ports  PCI bus for connecting master processor  register files replicated in each micro engine  on-chip scratch SRAM and I/O buffers  two sets of register files each micro engine  128 GPRs and 128 transfer registers  instruction set architecture  specified field for context switch  specified instruction for reading on- chip scratch SRAM

18. One application of Intel's IXP1200

19. Conclusions 1. what is a NP, why we need it and its features 2. benchmarks 3. processing unit architectures: CMP or SMP 4. fast handling I/O: task pipeline, coprocessor 5. memory architectures -- only a small part of a huge design space