Comment 18 for bug 879334

Jeff Ebert (jeffrey-ebertland) wrote :

There is an upstream bug here:

I have tried the latest mainstream kernel (3.5.0) using the instructions here:

I still see the high CPU load on the NFS server.

I then reversed the patch suggested in the above bug.

$ git show 9660439861aa8dbd5e2b8087f33e20760c2c9afc
commit 9660439861aa8dbd5e2b8087f33e20760c2c9afc
Author: Olga Kornievskaia <email address hidden>
Date: Tue Oct 21 14:13:47 2008 -0400

    svcrpc: take advantage of tcp autotuning

I also reversed the patch mentioned here manually, since I could not find the commit hash for it

Unfortunately, this patched version of 3.5.0 does not boot. I may have screwed up something else along the way, but I wanted to report this in case somebody has more time to experiment.

This particular patch looks like an ongoing problem for nfsd. It was reverted due to performance issues in 2009.

commit 7f4218354fe312b327af06c3d8c95ed5f214c8ca
Author: J. Bruce Fields <email address hidden>
Date: Wed May 27 18:51:06 2009 -0400

    nfsd: Revert "svcrpc: take advantage of tcp autotuning"

    This reverts commit 47a14ef1af48c696b214ac168f056ddc79793d0e "svcrpc:
    take advantage of tcp autotuning", which uncovered some further problems
    in the server rpc code, causing significant performance regressions in
    common cases.

    We will likely reinstate this patch after releasing 2.6.30 and applying
    some work on the underlying fixes to the problem (developed by Trond).

    Reported-by: Jeff Moyer <email address hidden>
    Cc: Olga Kornievskaia <email address hidden>
    Cc: Jim Rees <email address hidden>
    Cc: Trond Myklebust <email address hidden>
    Signed-off-by: J. Bruce Fields <email address hidden>

It was reintroduced in May 2011, commit a74d70b63f1a0230831bcca3145d85ae016f9d4c .

Hope this helps somebody...