Fix #20330
closedmsg : tcp backlog parameter increase
0%
Description
On a quite large cluster, during a cold restart phase, I noticed from tcp metrics some "ListenOverflows" and "ListenDrops" from /proc/net/netstat. On the other side, OSD's were failing to join MON's or peers (according to logs). Digging into, I tried to increase sysctl "net.core.somaxconn" kernel parameter (from some tutorials). Unfortunately, this parameter did not solved this issue. According to "ss -ln", nothing changed on the socket side.
Looking at linux documentation, I recompiled my packages modifying "128" hardcoded values from :
- https://github.com/ceph/ceph/blob/master/src/msg/simple/Accepter.cc : line 214
- https://github.com/ceph/ceph/blob/master/src/msg/async/PosixStack.cc : line 325
- https://github.com/ceph/ceph/blob/master/src/msg/async/rdma/RDMAServerSocketImpl.cc : line 59 (I do not use RDMA)
Values from "ss -ln" changed well this time and cluster reboot became smoother.
I think those values might evolve to something greater (> 1024) or even the maximum stock kernel accepts (65535) as it is capped by "somaxconn" value.
Another possibility would be to make it configurable.
Updated by Kefu Chai almost 7 years ago
- Status changed from New to Fix Under Review
- Assignee set to Haomai Wang
Updated by Sage Weil almost 7 years ago
- Status changed from Fix Under Review to Resolved
Updated by Greg Farnum about 5 years ago
- Project changed from Ceph to Messengers
- Category deleted (
msgr)