tuexen discussed how to mitigate blackhole detection problems
Long-term ideas include PMTUD, D-SACK-based probing, etc. Using real data introduces congestion control algorithm concerns.
tuexen found that RTO length has dropped since blackhole detection values were set. Those values are based on RTO length. Therefore, he proposes a short-term change of setting fixed values for blackhole detection timer (which could be higher than current values) to reduce false positives.
rscheff points out that they also care about probing back to increase MSS when the blackhole is resolved.
Discussion about whether D-SACK-based probing would work and be “simple enough” for a short-term solution.
Action item:
tuexen will commit change to allow an administrator to commit a fixed value. The used value will be max(current calculate value, fixed value). Therefore, a fixed value of 0 will maintain the current behavior.
rscheff will propose a patch with the D-SACK probing he suggests.
D23904: KMOD_TCPSTAT_INC vs. TCPSTAT_INC: tuexen will commit change to use stable KPI from modules.