Home | Overview

Fast AQM Scalable TCP

FAST 10GE Experiment

This is the planning page for the Jan 2003 experiment, and will be continually updated. Please email corrections/comments/suggestions.

Acknowledgments

Network:

• Network setup and other details

Status

• Test: Feb 25 - Mar 1
• w02gva is dedicated to 10Gbps test from Feb 25- Mar 1
• Router modules: 2 at CERN, 2 in Chicago, 1 in Sunnyvale
• NICs: 1 at CERN, 2 in Sunnyvale, 2 in Chicago

Experiments [Feb 25 - Mar 1]

1. For comparison purposes, all experiments should be done with
   • FAST kernel
   • Scalable TCP
   • HSTCP
   • TCP (optimized)
2. Periodic measurements to be taken to construct time traces (important to see how they behave in dynamic tests):
   • Throughput (iperf)
• RTT (in kernel)
• Loss (in kernel)
3. Single flow
• Goal: largest 1-flow throughput, largest bmps
• Sunnyvale -> CERN, 10GE NIC
• Sunnyvale -> Chicago, 10GE NIC (not available yet)
• Sunnyvale -> Amsterdam, GE end-to-end (path unavailable)
• Sunnyvale -> Tokyo via Abilene, GE end-to-end (availability?)
4. Multiple flows, with dynamic sharing
• Assume 3 (5) 10GE cards inSunnyvale and 3(5) in Chicago
• Goal:
• saturate OC192 with min #flows
• transient behavior in dynamic sharing
• fairness among FAST flows
• Sunnyvale -> Chicago
  • 3 (or 4) concurrent flows to saturate OC192.
• Sunnyvale -> Chicago
  • Dynamic sharing where #active flows varies, to observe how protocol reacts to congestion
  • Intervals between joining/departure of flows should be long enough to measure equilibrium properties (throughput, delay, loss, fairness in each period).
  • 3 (5 if available) 10GE flows starting one at a time successively at (say) 5min intervals, and then terminating one at a time at 5min intervals.
• Sunnyvale -> Chicago, CERN, Tokyo(?)
  • Similar to above, but with more flows to different destinations concurrently. This assumes Sylvain's Scenario 2 network setup.
  • Sunnyvale -> Chicago: 3(5) 10GE flows starting/departing successively
  • Sunnyvale -> CERN: 5 GE-flows starting/departing successively (This will saturate the DataTAG OC48 link)
  • Sunnyvale -> Amsterdam: 5 GE-flows starting/departing successively (path unavailable)
  • Sunnyvale -> Tokyo: 5 GE-flows starting/departing successively (availability?)
5. Fairness
   • N1 TCP flows compete with N2 FAST flows, where (N1, N2) = (0, 10), (1, 9), (2, 8), ..., (10, 0).
   • Sunnyvale -> Chicago (with 10GE-flows only)
   • Sunnyvale -> Chicago, CERN, Tokyo (with both 10GE-flows and GE-flows)
6. Disk transfers
   • See Julian Bunn's 3ware RAID arrays tests
• Sunnyvale -> Chicago, CERN(?)
• Disk -> Memory
• Memory -> Disk
• Disk -> Disk
7. QBSS
8. Benchmarking applications, e.g., PingER vs. netpipe
   vs. netperf vs. real app vs. ????
9. Unicast vs. multicast
   • single source to multiple sinks
   • single flow vs. "multiple flow/single source" vs.
     "multiple flow/multiple sources"
10. Different MTU
    • Standard MTU
    • 8K MTU
    • 9K MTU

History

• 1 Intel card at Sunnyvale and 1 at CERN around Feb 20. Intel card to arrive at Chicago later.
• Wu's team installed one 10GE card in a server in Chicago and started testing (one->many and many->one), Jan 16, 2003
• GSR will return to Cisco on Feb 25
• 1 Cisco 10GE modules arrived at SLAC Jan 27, 2003
• Two 10GE modules delivered to CERN on Jan 27
• Level3 OC192 between Sunnyvale & Chicago available till March
• DataTAG link between CERN & Chicago partitioned into 2 Gbps channels after March 3
• TeraGrid router T640 loan extended till end of Feb (Jan 16)
• 4 RAID at Sunnyvale and 4 at StarLight
• Amsterdam unreachable from Sunnyvale

Other Experiments

SC02 Caltech-SLAC Experiment: bmps, throughput traces, network setup, acknowledgments
SLAC webpage on SC2002: results, setup, equipment, etc

Les Cottrell's testing of FAST kernel (SC2002 version) at SLAC, Dec 17, 2002: throughput, CPU utilization, sharing.

Guided tour of background theory