I am using ceph mimic in a small test setup using the below configuration.

OS: ubuntu 18.04

1 node running (mon,mds,mgr) + 4 core cpu and 4GB RAM and 1 Gb lan

3 nodes each having 2 osd's, disks are 2TB + 2 core cpu and 4G RAM  and 1 Gb lan

1 node acting as cephfs client  + 2 core cpu and 4G RAM  and 1 Gb lan

configured cephfs_metadata_pool (3 replica) and cephfs_data_pool erasure 2+1.

When running a script doing multiple folders creation ceph started throwing error late IO due to high metadata workload.

once after folder creation complete PG's degraded and I am waiting for PG to complete recovery but my OSD's starting to crash due to OOM and restarting after some time.

Now my question is I can wait for recovery to complete but how do I stop OOM and OSD crash? basically want to know the way to control memory usage during recovery and make it stable.

I have also set very low PG metadata_pool 8 and data_pool 16.

I have already set "mon osd memory target to 1Gb" and I have set max-backfill from 1 to 8.

Attached msg from "kern.log" from one of the node and snippet of error msg in this mail.

---------error msg snippet ----------

-bash: fork: Cannot allocate memory

Sep 18 19:01:57 test-node1 kernel: [341246.765644] msgr-worker-0 invoked oom-killer: gfp_mask=0x14200ca(GFP_HIGHUSER_MOVABLE), nodemask=(null), order=0, oom_score_adj=0
Sep 18 19:02:00 test-node1 kernel: [341246.765645] msgr-worker-0 cpuset=/ mems_allowed=0
Sep 18 19:02:00 test-node1 kernel: [341246.765650] CPU: 1 PID: 1737 Comm: msgr-worker-0 Not tainted 4.15.0-45-generic #48-Ubuntu

Sep 18 19:02:02 test-node1 kernel: [341246.765833] Out of memory: Kill process 1727 (ceph-osd) score 489 or sacrifice child
Sep 18 19:02:03 test-node1 kernel: [341246.765919] Killed process 1727 (ceph-osd) total-vm:3483844kB, anon-rss:1992708kB, file-rss:0kB, shmem-rss:0kB
Sep 18 19:02:03 test-node1 kernel: [341246.899395] oom_reaper: reaped process 1727 (ceph-osd), now anon-rss:0kB, file-rss:0kB, shmem-rss:0kB
Sep 18 22:09:57 test-node1 kernel: [352529.433155] perf: interrupt took too long (4965 > 4938), lowering kernel.perf_event_max_sample_rate to 40250

regards

Amudhan