1. Supercharging PlanetLab A High Performance,Multi-Alpplication,Overlay Network Platform Reviewed by YoungSoo Lee CSL

2. Overview  PlanetLab has become a popular experimental platform.  But PlanetLab are subject to high latency, high delay jitter, poor performance. What is the Solution?

3. Solution General purpose server NP subsystems Supercharged PlanetLab Platform

4. SSP’s Objectives  Higher level of both IO performance and processing performance.  Reasonably straightforward for PlanetLab users take advantage of the capabilities  Require t.hat legacy PlanetLab applications run on the system without change.

5. System Overview  Modern NPs are not designed to be shared.  Application are divided into Fast Path and Slow Path

6. System Components  Line Card(LC) : LC forwards each arriving packet to the system component, and queues outgoing packets for transmission.  General Purpose Processing Engines(GPE).  Network Processing Engines(NPE)  Control Processor(CP)

7. Network Processor Issues  NP products have been developed for use in conventional router.  Micro-Engines(ME) : packet processing.  DRAM : packet buffer.  SRAM : implementing lookup table & linked list queues.  MP : overall system control

8. Network Processor Issues  NP provides a different mechanism for coping with the memory latency gap, hardware multithreading.  Operate in a round-robin fashion.

9. Sharing the NPE  Develop software for the NPE that allows it to be shared by the fast path segments of many different slices.  Process 8 packets concurrently using the hardware thread contexts.

10. Sharing the NPE  Rx : packet received from switch are copied to DRAM & passes a pointer main packet processing pipeline.  Substr. : determines which slice the packet belongs & strips outer header form the packet.  Parse : preconfigured set of Code Option & form lookup key.

11. Sharing the NPE  Lookup : provides generic lookup capability.  Hdr Format : makes necessary changes to the slice-specific packet header.  Queue Manager : implements a configurable collection of queues.  Tx : forwards the packet to the output.

12. Enhancing GPE Performance  Boost performance of the GPEs.  1. Use higher performance hardware configurations than usual for PlanetLab.  2. Improve the latency of PlanetLab applications.  Changing coarse-grained scheduling paradigm

13. Enhancing GPE Performance Token-based scheduling Max. number of token : 100 Min. number of token : 50 N vServer pre-empted 100(N-1) ms at a time. Token-based scheduling The Min. & Max. token allocation were set same value Varied from 2 to 16

14. Enhancing GPE Performance

15. Overall System

16. Slice Configuration  ① : CP obtains slice configuration data using the standard PlanetLab mechanism of periodically polling the PLC  ② : Slices are assigned to one of GPE by GNM.  ③ : Corresponding entry is made in a local copy of my PLC.  ④ : The LNM on each of the GPEs periodically poll myPLC to obtain new slice configurations. ① ② ③ ④

17. Port Assignment  ① : LRM reservation request th the GRM.  ② : If the requested port number is available, it makes the appropriate assignment to LRM.  ③ : GRM configures the Line Card so that LC will forward to the right GPE. ① ②③

18. NPE Assignment  ① : LRM forwards the request to the GRM.  ② : GRM selects the most appropriate NPE to host the slice and returns its id to the LRM  ③ : LRM then interacts with the MP. ① ② ③

19. Evaluation  Implement two different application.  IPv4 & Internet Indirection Infrastructure.

20. IPv4-Fair Queueing Mechanism

22. IPv4 - Throughput

23. IPv4 - Latency

24. Internet Indirection Infrastructure Performing throughput and latency tests similar those we did for IPv4 application. Achieve 30~40% higher result.

25. IPv4 & III- Throughput

26. III- GPE vs NPE