You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Doing some prototyping using Quiche with WebTransport+BBR2, I found the datagram queue appeared to be ignored when assessing whether the app is limiting the bitrate:
In my test, where the pipe is filled with datagrams, BBR constantly marks datapoints as app limited and never gets out of STARTUP phase.
Simply considering a non-empty datagram queue as willingness to write allows BBR to get into PROBE_BW, although I'm sill seeing poor behavior, where the RTT increases quite dramatically from ~40ms (min RTT) to sustained ~200ms when we hit a bitrate close to congestion, but I haven't looked into what's causing it (no draining on latency increase?)
The text was updated successfully, but these errors were encountered:
Included changes:
- c7b993eb750e60c307e82f75763600d9c06a6de1 Merge pull request google#28 from vasilvv/master by Victor Vasiliev <vasilvv@google.com>
- da778f0a1cf158d3a6fbfa573d603efc4ead4b7e Add a missing deps in the BUILD file by Victor Vasiliev <vasilvv@google.com>
- 0146ddf01a02c32761e64c2e0ef35f3e56c64d9a Merge pull request google#27 from vasilvv/master by Victor Vasiliev <vasilvv@google.com>
- 764cca256a881c6189245d967fa81adf5a0676c6 Replace "const std::string&" with absl::string_view by Victor Vasiliev <vasilvv@google.com>
PiperOrigin-RevId: 441304196
Doing some prototyping using Quiche with WebTransport+BBR2, I found the datagram queue appeared to be ignored when assessing whether the app is limiting the bitrate:
quiche/quiche/quic/core/quic_session.h
Line 979 in 9b45510
quiche/quiche/quic/core/quic_session.cc
Line 711 in 9b45510
quiche/quiche/quic/core/quic_connection.cc
Line 5277 in 9b45510
In my test, where the pipe is filled with datagrams, BBR constantly marks datapoints as app limited and never gets out of STARTUP phase.
Simply considering a non-empty datagram queue as willingness to write allows BBR to get into PROBE_BW, although I'm sill seeing poor behavior, where the RTT increases quite dramatically from ~40ms (min RTT) to sustained ~200ms when we hit a bitrate close to congestion, but I haven't looked into what's causing it (no draining on latency increase?)
The text was updated successfully, but these errors were encountered: