[OpenAFS-devel] Re: Performance single proc. vers. SMP

[Hartmut Reuter]

Some more results and details:

My laptop runs SuSE-7.3 with kernel 2.4.10. I use a ramdisk-cache of 16
MB in both cases (SMP and single processor kernel).

To make sure it is really the AFS code which makes things slow I mounted
AFS via NFS on /mnt (exported by a userland-nfsserver from a fast SMP
machine).
I get a read performance of 4.5 MB/s which is only slightly lower than
AFS access with the SMP kernel, but more than twice what I get reading
directly from AFS with the single processor kernel.

My feeling is that earlier versions of OpenAFS were faster with the
single processor kernel. I must do some more tests to verify this.

Any comments?

Hartmut

Hartmut Reuter wrote:
> 
> I made a strange observation:
> 
> Running openafs (from cvs) on my laptop I got rather bad read
> performance on large files: about 1.3 MB/s.
> 
> Then I built an SMP-kernel, rebooted with that one and built the the
> libafs for SMP. With this one I can read with 5 MB/s.
> 
> I also made fstraces in both cases.
> 
> It turns out that the trace lines inside afs_UFSCacheFetchProc() which
> appear whenever he wakes up waiters for tdc->f.validPos come rather
> continously in the SMP case, but in the single processor case appear
> some times long delays.
> 
> This lets me think of a problem in the rx-layer because in this loop
> nothing else happens than rx_Read() and afs_osi_Write(). I suppose
> afs_osi_write() is fast and could only be faster in the single-processor
> kernel.
> 
> Any ideas?
> 
> Hartmut
> -----------------------------------------------------------------
> Hartmut Reuter                           e-mail reuter@rzg.mpg.de
>                                            phone +49-89-3299-1328
> RZG (Rechenzentrum Garching)               fax   +49-89-3299-1301
> Computing Center of the Max-Planck-Gesellschaft (MPG) and the
> Institut fuer Plasmaphysik (IPP)
> -----------------------------------------------------------------

-- 
-----------------------------------------------------------------
Hartmut Reuter                           e-mail reuter@rzg.mpg.de
					   phone +49-89-3299-1328
RZG (Rechenzentrum Garching)               fax   +49-89-3299-1301 
Computing Center of the Max-Planck-Gesellschaft (MPG) and the
Institut fuer Plasmaphysik (IPP)
-----------------------------------------------------------------