1. Supercharging PlanetLab : a high performance, Multi-Application, Overlay Network Platform Written by Jon Turner and 11 fellows. Presented by Benjamin Chervet 8 April 2009 1

2. Content : Context Motivation How to supercharge PlanetLab ? Evaluation Summary 2

3. Context : Overlay Network ? Computer Network built on Top of another Examples – P2P Network over the Internet – Dial Up Internet over Telephone Network – PlanetLab over the Internet 3

4. Context : PlanetLab Group of computers over the Internet Used as a testbed for research and deployment Shared overlay infrastructures. 4

5. Motivations : PlanetLab limitations Applications run as user-space processes in virtual machines. – Limited throughput – High, unpredictable latency => What is the solution ? 5

6. Supercharging: How works a node right now ? 6

7. Objectives of a supercharge nodes Better IO and processing performances Not too hard for users to take advantage of the modifications. Current PlanetLab should run on the new nodes. 7

8. Supercharged node 8

9. Architecture of a supercharged node Conventional server, coordinates system components and synchronizes with PlanetLab Server blade supporting standard PlanetLab environment. Forward the packets to correct PlanetLab environments Network processor supporting applications fast-paths Blade containing data switch and control switch. 9

10. What is a Network processor ? Use of Network processor technology – A kind of processor specially designed to handle network operations. – Programmable – Typically used in Routers, switchs and firewalls. 10

11. Architecture of a NP devices 16 Multi threaded (Mes) – Fast nearest neighbor connections for pipelined apps. 3 SDRAM and 4 SRAM channels Management Processor for control 11

12. NP devices How a NP takes advantages from multi threading ? – Cope with the memory latency gap. – Round robin 12

13. System Control Divided in two paths: Pink : Normal Path Blue : Fast Path 13

14. Using the fast path Process 8 packets concurrently. Fast path shared by slices more efficiently than the old fashion nodes. 14

15. GPE improvements Use better performance hardware Change the scheduling of applications to improve the latency. – Number of token reduced. 15

16. Evaluation of the improvements Implementation of real world applications. – IPv4 arrive/depart in UDP tunnels. – Internet Indirection Infrastructure (i3) 16

17. I3 – Packets contain triggers matched to IP addresses – No match at local node results in Chord forwarding 17

18. IPv4 Throughput Comparison 18

19. IPv4 Latency Comparison 19

20. I3 Latency Comparaison 20

21. IPv4/i3 Fast Path throughput comparison 21

22. Summary Supercharged nodes remove the limitations of PlanetLab Nodes Easy to deploy and adapt to existing structures. Future work: – Automatic NPE code verification – Flexible NPE implementations 22