6/23 Discord Meeting

[Nick Banks (nibanks)]

Agenda

1. Discussion of Implement BBR congestion control algorithm #2071 between Wizmann (@Wizmann) and Matt Olson (@maolson-msft) (if both available).
2. Discussion of recent XDP latency work.
   a. Improvements to WPA tooling (w/ demo).
   b. Functional improvements in the code.

[Nick Banks (nibanks)]

Notes

(Thanks Daniel M. Havey (@dahavey) as usually for recording these!)

• Goal
  • See how a single stream exists relative to other streams
• Is the way we batch the rcvs the best for latency?
  • We are trying out code to spread the rcvs for comparison
• This is a question of batching vs pacing.
  • Pacing lets the app serialize the work better reducing queues
• The more streams you add the more queuing pressure you see at different layers
• How can improve latency and reduce execution time?
  • Break down and analyze log files
• Built additional tooling to join both log files
• Currently only supports Windows logs
  • Working on tooling to support Linux logs
• Logic to improve rcv flushing
• Improvements in asynch completion path
  • Upcall to the app rcv, downcall to us rcv_complete
• Our code was queuing. (Fixed)
  • Lot of work at the XDP/MsQuic interface
• Planning for QUIC now
  • Taking feature requests from community
• Is there a plan for a JAVA implementation of MsQuic
  • We have explored exposing a JAVA API
• Hyper_X_pro to look into this: Support Java Runtime #1141
• Questions about Linux XDP support:
  • Is this on our radar at all?
• Yes, definitely.
  • Perf on Linux is good for us, but we need to be strategic with our resources
• Perf is always a priority for us
  • If perf is lower than it should be in any code path then we need to know about and understand it.
• Also there are plenty of features that we would like to support
  • 0RTT
  • Client migration
• Did we publish perf numbers with encryption disabled?
  • Yes, it was 21 Gbps
• How do you grab a cpu trace on linux?
  • How do you dig into it?
  • Can we get information as detailed as WPA
  • We need to be able to ingest LTTng format.
  • Typical tool for perf investigation on Linux ValGrind.
• You can see stack traces similar to WPA or Xperf.
• We could figure out a lot starting with this tool.
  • Azure profiler -> gives flame graphs