1. Communications in ISTORE Dan Hettena

2. Communication Goals Goals: Fault tolerance through redundancy Tolerate any single hardware failure High bandwidth with commodity hardware Use redundancy for extra bandwidth Lower latency alternative For latency-sensitive apps, such as Titanium Provide Active Messages interface

3. Outline ISTORE Network Hardware IP Communication Actives Messages Communication

4. ISTORE Network Hardware Components 64 ISTORE Bricks, each with: Pentium II 266MHz IBM 10kRPM SCSI disk Can sometimes be read faster than 30MB/s 4 100Mbps ethernet interfaces Intel EtherExpress Pro/100 (82557/8) Total bandwidth = 4*100Mbps = 40MB/s

5. ISTORE Networking Hardware Components (continued) 14 “Little” Routing Switches PacketEngine/Alcatel PowerRail 1000 20 100Mbps interfaces (copper) 2 1Gbps interfaces (fiber) 2 “Big” Routing Switches PacketEngine/Alcatel PowerRail 5200 More-than-enough 1Gbps interfaces (fiber)

6. ISTORE Networking Hardware Routes between bricks

7. ISTORE Networking Hardware Routes between bricks (continued) Short routes Only if connected to the same “little” switches No need to go through a “big” switch 2 hops

8. ISTORE Networking Hardware Routes between bricks (continued) Long routes Must choose a big switch 4 hops

9. ISTORE Networking Hardware Performance observations Switches are store-and-forward Ethernet packets are all at least 60 bytes 0-padded by sender if necessary Time per 100Mbps copper hop is 15  s + (10ns/bit)(size – 60 bytes)

10. ISTORE Networking Hardware Future work Plug in the wires

11. IP Communication Goals Stripe packets across all 4 interfaces 4x increase in available TCP/UDP bandwidth Automatically handle link and router faults Transparent to TCP/UDP applications Transparent backward-compatibility with hosts that do not support striping

12. IP Communication Nested outline Previous work Kernel driver overview Providing fault tolerance Providing backward compatibility Making sure it scales

13. IP Communication Previous work Linux bonding driver (net/drivers/bonding.c) Generic driver to “bond” links Ignores faults Does not prevent packet reordering Only supports one remote host “An Architecture for Packet-Striping Protocols” (Hari, Varghese, Parulkar)

14. IP Communication Kernel striping driver Cooperates with ethernet driver Use special MAC addresses 49:53:54:4F: : Easy to determine if host supports striping Store striping information in headers Link status Reordering data

15. IP Communication Fault tolerance User process periodically tests links Notifies striping driver Striping driver will not use broken links Need to detect performance faults, too Backward compatibility Same IP address for both modes

16. IP Communication Scales automatically, unless packets arrive out of order Possible with multiple routes (e.g. striping) TCP relies on a consistent round-trip time Packet reordering confuses TCP Result is unnecessary retransmissions This will be an issue in ISTORE

17. IP Communication Scaling (continued) Need to reorder packets before handing them up to IP Solution (almost implemented) Clever use of queuing algorithms on sender and receiver Makes it unlikely that the receiver will dequeue packets out of order This is previous work (Hari et al)

18. IP Communication Future Work Complete reordering support Automatic detection of node ID Automatic detection of incorrect wiring Automatic configuration of the switches By the Diagnostic Processors

19. Active Messages Support Goals Support latency-sensitive apps Titanium, for example This is not a primary goal for ISTORE As reflected in the networking hardware Non-goals Transparency Support for malicious users

20. Active Messages Support Problem: kernel ruins latency Protocol stacks (UDP, TCP) are slow User-kernel interaction is slow Solution: remove kernel from critical path Previous work: U-Net User level ethernet driver Cooperates with kernel driver Only accessible to trusted users

21. Active Messages Support Custom AM implementation By Dan Bonachea (not me) Based on HPAM But also supports big bulk transfers Supports ISTORE user-level networking Automatically gets a new network card (if a link dies) Also supports UDP

22. Active Messages Support Performance comparison Using a ping program (client of AM) Using short route (four hops) Ethernet latency = 4*(17  s)=68  s UDP mean round-trip time is 160  s User-level ethernet mean is 80  s Includes check-summing and AM overhead

23. Active Messages Support Future work Compile Titanium for ISTORE and see what happens.

24. Conclusions Kernel hacking is fun And my talking is done.