Fwd: [Gluster-devel] Performance

[Jerker Nyberg]

Hi,

I've done some testing on old hardware (5 nodes with 1 GHz Athlon, 512 MB 
RAM, 40 GB IDE-disks, 100 Mbit/s ethernet). Both NFS and GlusterFS fill 
the ethernet when using large blocksizes. For small blocksizes however 
GlusterFS is running out of CPU... Is there any good way to minimize the 
CPU usage on the GlusterFS client and to increase small blocksize 
performance? I'm trying different combinations with the translators to see 
if I get slightly better configuration (but after removing all of them 
bonnie++ take ages to complete.) :)

bonnie++ 1.03:
                  Seq Output-----  Seq Input-----
                  Per Chr    CPU   Per Chr    CPU
    local disk:   17324 K/s  94%   18154 K/s  89%
    GlusterFS:     4305 K/s  21%    7195 K/s  36%
    NFS:           6197 K/s  29%   11253 K/s  56%

Well, NFS seem to be more efficient on this old hardware than GlusterFS...

One of the reasons that I find GlusterFS interesting is the possibility to 
increase the number of IO per second when using many drives in parallell. 
That has been the major performance bottleneck for me when running 
mail/webhosting. More drives help, and adding more servers instad of 
buying large RAIDs would be neat. In a way, yes it seem to scale, but as 
far as I understand at the cost that every node will spend more CPU 
accessing the file system. Perhaps not a problem on modern hardware?

I also tried some bittorrent-seeding from GlusterFS. For 
bittorrent-clients not using much read-ahead cache this is normally 
limited by the disk seek-time (IO/s). I was hoping that utilizing remote 
disks with a total of five times the IO/s than the local disk would 
increase the seeding rate, but instead GlusterFS and the bittorrent is 
just competing for CPU. Well well, for bittorrent the problem seem to be 
solved with readahead cache in the modern clients anyway, increasing the 
RAM cache hit-rate a lot.

All bonnie output here:
http://www.update.uu.se/~jerker/tmp/glusterfs.bonnie.txt

Regards,
Jerker Nyberg.


On Mon, 15 Oct 2007, Steffen Grunewald wrote:

> What's not so beautiful is that the first dd (always nfs) does
> include staging of the file from the input media into buffer
> cache (/dev/zero means: filling memory with zero bytes, which
> certainly is faster than reading from a physical disk).
> I would have repeated the write tests to see whether ordering
> is important:
> - nfs write
> - glusterfs write
> - nfs write again
> - glusterfs write again
> Buffers are often able to fool the benchmarker.
>
> Also some information about your machine is missing - but I suppose 1GB
> would easily fit into main memory. What about *several* GBs to effectively
> trash the page cache?
>
> Cheers,
> Steffen, always doubtful when it comes to benchmarks
>
> -- 
> Steffen Grunewald * MPI Grav.Phys.(AEI) * Am Mühlenberg 1, D-14476 Potsdam
> Cluster Admin * http://pandora.aei.mpg.de/merlin/ * http://www.aei.mpg.de/
> * e-mail: steffen.grunewald(*)aei.mpg.de * +49-331-567-{fon:7233,fax:7298}
> No Word/PPT mails - http://www.gnu.org/philosophy/no-word-attachments.html
>
>
> _______________________________________________
> Gluster-devel mailing list
> Gluster-devel at nongnu.org
> http://lists.nongnu.org/mailman/listinfo/gluster-devel
>