Description of problem: ======================= After rebooting the nodes which are hosting 400 volumes bricks, failed to start some of the volume bricks. Errors in glutserd logs: ======================= [2016-05-13 08:16:04.924247] E [socket.c:2393:socket_connect_finish] 0-glusterfs: connection to 10.70.36.45:24007 failed (Connection timed out) [2016-05-13 08:16:05.128728] E [glusterfsd-mgmt.c:1907:mgmt_rpc_notify] 0-glusterfsd-mgmt: failed to connect with remote-host: rhs-client21.lab.eng.blr.redhat.com (Transport endpoint is not connected) [2016-05-13 08:16:05.340730] I [glusterfsd-mgmt.c:1913:mgmt_rpc_notify] 0-glusterfsd-mgmt: Exhausted all volfile Version-Release number of selected component (if applicable): ============================================================= glusterfs-3.7.9-4. How reproducible: ================= Always Steps to Reproduce: =================== 1. Have two RHGS node with 16 GB RAM each. 2. Create 400 1*2 volumes using both the nodes and start all the volumes. 3. Reboot the nodes and check all volume bricks are running. Actual results: =============== Bricks not starting after node reboot. Expected results: ================= Bricks should start after rebooting of nodes. Additional info:
This looks like that GlusterD is not able to communicate with bricks due to lack of multi threaded e-poll support in GlusterD. [2016-05-13 08:16:04.924247] E [socket.c:2393:socket_connect_finish] 0-glusterfs: connection to 10.70.36.45:24007 failed (Connection timed out) [2016-05-13 08:16:05.128728] E [glusterfsd-mgmt.c:1907:mgmt_rpc_notify] 0-glusterfsd-mgmt: failed to connect with remote-host: rhs-client21.lab.eng.blr.redhat.com (Transport endpoint is not connected) [2016-05-13 08:16:05.340730] I [glusterfsd-mgmt.c:1913:mgmt_rpc_notify] 0-glusterfsd-mgmt: Exhausted all volfile >From the above log (especially the first one) this indicates that the same brick process failed to connect to glusterd and the connection got timed out. This can happen in a situation where there is lot of back pressure on the other side. Since GlusterD is limited to a single threaded e-poll communication with the brick processes happen over a single path and hence while glusterd tried to start 400 odd brick processes there were 400 RPC connections to handle and that's why few of the brick process got to hear from GlusterD and they came up but others did not. With out a brick multiplexing feature in place, scaling volumes is going to be always a challenge with different set of problems. Moving it to 3.2.
I tried to enable MT-epoll on a set up of 4 nodes, 400 volumes with bricks spanning over all the nodes. After rebooting a node, all the gluster brick processes didn't come up and same error message was seen in few brick log file. So on a nutshell, MT-epoll is not going to solve this scalability issue. Its the big lock which is causing the threads to block and time out.
Surprisingly, big lock is not a culprit here. Its the pmap_signin from the brick processes which was consuming lot of glusterd's bandwidth and a code walk through revealed that we were doing an unnecessary address resolution which was not needed. Applying the fix solves this problem and I could see that on rebooting, glusterd is able to bring up all the brick processes.
http://review.gluster.org/#/c/14849/ posted for review.
Upstream mainline : http://review.gluster.org/14849 Upstream 3.8 : http://review.gluster.org/14860 And the fix is available in rhgs-3.2.0 as part of rebase to GlusterFS 3.8.4.
Verified this bug using the build glusterfs-3.8.4-3, Reported issue is not seen any more. Created 500 1*2 volumes using two nodes having 32G RAM and done stop and start of volumes, it worked well and done node reboots to check the issue reported,all the volume bricks were up. Moving to verified state.
LGTM :)
Since the problem described in this bug report should be resolved in a recent advisory, it has been closed with a resolution of ERRATA. For information on the advisory, and where to find the updated files, follow the link below. If the solution does not work for you, open a new bug report. https://rhn.redhat.com/errata/RHSA-2017-0486.html