[OpenAFS-devel] read performance 1.5.74 Linux

[Hartmut Reuter]

Simon Wilkinson schrieb:

> If you can reproduce this using a command line tool such as afsio, then we
> could trying and track this down using git bisect - that's probably going to
> be the best route. Testing with afsio would also let us implicate the RX
> library, rather than the cache manager, as being the source of the problem.

afsio seems to be visibly faster in the 1.4.12-osd version than in the
1.5.74-osd version. You are right, they differ nearly only in the rx-library.
However, the difference is not a factor 10 but a factor 1.5, only.

I reach 18 MB/s for read with the 1.5.74 afsio and 27 MB/s with that of 1.4.12.

In both cases I read an 1gb file on my external usb-drive used as disk for the
rxosd.

When I read the same file through the 1.5.74-osd client with vicep-access
enabled I reach 35 MB/s. But then most of the rx-traffic is avoided. So these 35
MB/s are about the speed of the disk.

The rx-layer of userland programs is not identical with that in the kernel. So
the lots of packet reclaims we see with the 1.5.74 client and subsequent packet
losses which cause timeouts and resends seem to be the main bottleneck in the
kernel client. Obviously they do not appear with afsio whatever version of rx it
was linked against (seen with rxdebug -peer -long to the rxosd).

Hartmut

> 
> Cheers,
> 
> Simon.
> 
> _______________________________________________ OpenAFS-devel mailing list 
> OpenAFS-devel@openafs.org 
> https://lists.openafs.org/mailman/listinfo/openafs-devel


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