1

I recently set up a new nfs server, a proxmox6 VM with ceph as storage backend. The server itself is fast making read/write of > 300MBytes/sec possible on the local filesystems that are exported via NFS. NFS access however is very slow when writing is concerned: writing files, removing files etc. I created a test directory with 1.5 MB in size and ~300 small files in it:

    nfsserver# mount|grep krienke
    nfsserver:/user2/krienke on /home/krienke type nfs4 (rw,noatime,vers=4.2,rsize=1048576,wsize=1048576,namlen=255,hard,proto=tcp,timeo=600,retrans=2,sec=krb5,clientaddr=x.y.z.a,local_lock=none,addr=x.y.x.z.b)

    # Copy via SSH to NFS directory on the nfsserver which NFS-automounts 
    # /export/user2/krienke on /home/krienke if accessed
    client$ time scp -r etc nfsserver:/home/krienke/  
    real    2m34,598s

    # No NFS involved:
    client$ time scp -r etc nfsserver:/export/user2/krienke/
    real     0m3,932s

You see the difference in execution time between non NFS access and access when NFS is involved. The same happens when I do NFS from a client machine. The NFS server is a SLES15SP1 machine with 10GB ethernet. The client has a 1Gbit interface. On the servers eth0 interface I see a view dropped packages which are about 0.003% of the totally xferred packages, which does not seem to be something to worry about. When copying the test-directory "etc" on the server itself not touching nfs the copy command is done immediately after having pressed ENTER. So for me it seems to a NFS related problem. I can reproduce this problem with both NFS V3 and NFS V4 access(the server exports both). Unlike writing, reading from NFS is fine.

One thing I see in nfsstat -s -v3 a quite high getattr value:

Server nfs v3:
null             getattr          setattr          lookup           access           
4757      0%     9660323  84%     268638    2%     163260    1%     284994    2%     
readlink         read             write            create           mkdir            
396       0%     87230     0%     223067    1%     58621     0%     1799      0%     
symlink          mknod            remove           rmdir            rename           
25        0%     0         0%     40817     0%     903       0%     23826     0%     
link             readdir          readdirplus      fsstat           fsinfo           
0         0%     22242     0%     249358    2%     314048    2%     968       0%     
pathconf         commit
484       0%     33792     0%

My problem is up to now I did not find any kind of misconfiguration or such and I need some ideas where I could go searching for the source of this problem?

Thanks Rainer

rainer042
  • 33
  • 4
  • you can try to identify slow requests on NFS by running tshark on the nfsserver host: `tshark -f "port 2049" -Y "rpc.time > 0.001" -i any -T fields -e frame.number -e frame.time_delta -e ip.src -e ip.dst -e col.Protocol -e rpc.time -e col.Info` – kofemann May 12 '20 at 11:18
  • Thanks for your answer.I will try this. At the moment I switches to async export. Such shares work very fast, however with the potential problem of dataloss. – rainer042 May 16 '20 at 09:35

0 Answers0