Hi,
At UZH we are currently evaluating cephfs as a distributed file system
for the scratch space of an HPC installation. Some slow down of the
metadata operations seems to occur under certain circumstances. In
particular, commands issued after some big file deletion could take
several seconds.
Example:
dd bs=$((1024*1024*128)) count=2048 if=/dev/zero of=./dd-test
274877906944 bytes (275 GB, 256 GiB) copied, 224.798 s, 1.2 GB/s
dd bs=$((1024*1024*128)) count=2048 if=./dd-test of=./dd-test2
274877906944 bytes (275 GB, 256 GiB) copied, 1228.87 s, 224 MB/s
ls; time rm dd-test2 ; time ls
dd-test dd-test2
real 0m0.004s
user 0m0.000s
sys 0m0.000s
dd-test
real 0m8.795s
user 0m0.000s
sys 0m0.000s
Additionally, the time it takes to complete the "ls" command appears to
be proportional to the size of the deleted file. The issue described
above is not limited to "ls" but extends to other commands:
ls ; time rm dd-test2 ; time du -hs ./*
dd-test dd-test2
real 0m0.003s
user 0m0.000s
sys 0m0.000s
128G ./dd-test
real 0m9.974s
user 0m0.000s
sys 0m0.000s
What might be causing this behavior and eventually how could we improve it?
Setup:
- ceph version: 10.2.9, OS: Ubuntu 16.04, kernel: 4.8.0-58-generic,
- 3 monitors,
- 1 mds,
- 3 storage nodes with 24 X 4TB disks on each node: 1 OSD/disk (72 OSDs
in total). 4TB disks are used for the cephfs_data pool. Journaling is on
SSDs,
- we installed an 400GB NVMe disk on each storage node and aggregated
the tree disks in crush rule. cephfs_metadata pool was then created
using that rule and therefore is hosted on the NVMes. Journaling and
data are on the same partition here.
So far we are using the default ceph configuration settings.
Clients are mounting the file system with the kernel driver using the
following options (again default):
"rw,noatime,name=admin,secret=<hidden>,acl,_netdev".
Thank you in advance for the help.
Cheers,
Tyanko
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
