[OpenAFS] low write/read performance with 1.8.x 1.9.0 client

[Benjamin Kaduk]

On Tue, Oct 13, 2020 at 09:42:14AM -0400, Jeffrey E Altman wrote:
> On 10/13/2020 9:28 AM, Giovanni Bracco (giovanni.bracco@enea.it) wrote:
> > I have seen that the first release of OpenAFS 1.9.0 is out and so I
> > thought that it was time to try at least 1.8.x and also 1.9 on our
> > production Linux x86-64 nodes, where we have used 1.6.x up to now.
> > 
> > Our AFS cell has file servers with OpenAFS 1.6.x and while over WAN the
> > performance can be rather poor, due to the well known rx latency
> > problem, in the LAN we have values between 70 e 80 MB/s both for read
> > and write. The clients are CentOS 6.x and 7.x all with OpenAFS 1.6.x.
> > 
> > So we have tried OpenAFS 1.8.6 clients on some production nodes (from
> > CentOS 7.3 to 7.8) and the performance on LAN have been poor, 15 MB/s
> > while the same nodes with 1.6.x clients has the normal performance.
> > 
> > Then we have installed a test AFS cell with OpenAFS 1.8.6 but no change
> > and at that point I we have checked also on user desktops using ubuntu.
> > 1.8.4 and the performance are as low as for production nodes with 1.8.x
> > release.
> > No improvement going to 1.9.0 (both server and clients) either.
> > 
> > Any suggestion?
> > Are we missing something important?
> > No special tuning in our installation.
> > 
> > Giovanni
> 
> I suspect the observed performance difference is due to:
> 
> commit 6d59b7c4b4b712160a6d60491c95c111bb831fbb
> Author: Benjamin Kaduk <kaduk@mit.edu>
> Date:   Sun Jul 30 20:57:05 2017 -0500
> 
>   Default to crypt mode for unix clients

I agree, though I probably would have copied the relevant text from the
1.8.0 release notes:

  All Client Platforms

    * Use rxkad_crypt by default for connections to fileservers.  This matches
      the existing behavior of the Windows client and has been applied by
      the distribution packaging on many platforms already.


-Ben