6/9 Discord Meeting

[nibanks]

Note: I plan to start creating a discussion item before every Discord meeting to put together an agenda (and allow others to add to it), and then attach any notes from the discussion after the meeting.

Agenda

• Discuss Active PRs:
  • Implement BBR congestion control algorithm #2071
  • Ensure Platform globals are freed on error #2685
  • Support multiple expected certificate validation results in tests #2782
  • Add support for Schannel in-proc certificate API. #2785
  • The various Get/Set Param PRs by Daiki
• Discuss In-Progress Work:
  • Advanced client certificate support #2573
  • Improve Schannel certificate retrieval speed #2750
  • More Comprehensive Certificate Testing #1172
• Discuss Important Uncommitted Work:
  • Support Dynamic Client Certificate #1249
  • Support Callback Negotiation Between Matching ALPNs #2418
  • MsQuicStreamSend may return QUIC_STATUS_ABORTED before indicating corresponding event. #2727
• Discuss Other Issues:
  • Log request start and end time in secnetperf #2800
  • ThroughputUp_LinuxSharedEC_x64_openssl_Default_Loopback is Failing #2794
• Discuss Ongoing LoLa Perf Work

[dahavey]

Meeting Notes:
• Adding API changes
• Peer Certification
o Serialized cert store - need to use CAPI
o Peer Certificate CAPI Format
• What about portable certificates?
o Maybe: Fast Schannel(?)
o Call it InProc Peer Certificate - maps to the Schannel PnProc flag. It maps to the Schannel InProc.
o Need more tests so that we can make sure that we can execute the code and it doesn't crash
• Daiki - 5 PRs out, please start getting them merged. As they are ready flag them for review.
• Wizman/Tony - busy making slides, can make some time this month for the things we want to do.
• BBR - Goal is to get the PR merged by the end of the month
o Wizman, yes will work with Matt.
• …Lost connection…
o Just started to go through the BBRv2 code
• It is double the size of the code - much more complex.
o This will take some time
o Must be labeled as preview
• Opt-in only
o BBR for private networks only.
• Matt and Wizman to work together to get this checked in by EoM
o Might need more time.
o Check in what you have now.
• It looks good, but needs some documentation.
 More comments
• PRs:
o Advanced client cert
• Need to test revocation flags.
• Everything else works.
• Close with comment - We have partial validation from HTTP team.
o Improve Schannel Certificate retrieval speed
• Need to merge this PR
o More comprehensive cert testing
• Untrusted certificate tests added(?)
 Please go through PR, check whatever boxes that can be checked. Add more if needed
o Support Dynamic Certificate
• Not sure if this is possible in Schannel or OpenSSL
• Either the TLS libs support it or we would have to reprocess it.
 What happens if the cert context is different?
• Lots of complexity
• Ask Thomas or Mathew to try to get resolution on what is possible
• Support callback negotiation between matching ALPNs
o Uncertain of priority right now.
• We are understaffed and don't want to commit without .NET buy in.
o ALPN selection would have to be done in the handshake
o Will this block .NET 7 from shipping?
• MsQuicStreamSend
o Want guarantee of serialization
• Can we change things around to give them the serialization
• Or can we just give them an API to query?
o Will they consume this if we build it?
o They call send while our thread is aborting stuff and they get the failed send call before the abort notification.
o The send call would have to block or succeed until the abort call
• The cleanup would have to happen after this. Right now it happens before.
• Improving the dashboard
o Fixed bug with shared exec context Linux
• Wasn't really using shared exec context
o Latency for shared exec context on Linux is really low. Better than Windows Kernel
• Not better than XDP
• Had meeting with UC Louvain
o They are interested in DPDK, XDP, look into an interop runner for perf.
o This would be very interesting. MsQuic is the fastest.
• Are we the fastest by a lot? By a little?
• Meeting with some academics out of Georgia Tech
o Will see if they want to publish a paper
• Would like to get Linux perf on par with Windows Kernel perf.
o Linux perf is terribly bi-modal.
• The shared EC flattens out the bi-modality
o It might be cross-NUMAing
• If it goes cross-NUMA then the perf is low. Otherwise the perf is good.
• Interested in Linux, but don't have a lot of scenarios
• MacOS is missing all of the APIs needed to be fast.