[Gluster-users] Write-behind translator and network traffic.

Cory Meyer cory.meyer at gmail.com
Fri Feb 6 23:14:27 UTC 2009


I've got the same config (Client side AFRx2 w/write-behind) that I've been
trying to optimize for small file writes and optimization of write-behind is
thus very important in my own case.    Using DD and Postmark for
benchmarking between GlusterFS 1.3.12, 2.0.0rc1 and
glusterfs--mainline--3.0--patch-888  though have seen no performance
difference with and without AFR and it's tweaks.

Here is a link to my test config..
http://gluster.pastebin.com/m46cb6a70

Benchmarking with DD and various block sizes produces a wide range of
results.. Optimally a block size of 1MB produces results of 70MB/sec.

4k blocks 4.3 MB/s
8k blocks 8.3 MB/s
16k blocks 15.4 MB/s

Is there a better way to tune write-behind since after a month of tinkering
I still don't see any performance increase in write-behind?

Thanks,

 Cory


On Fri, Feb 6, 2009 at 3:09 AM, Rash <rash at konto.pl> wrote:

> Hey!
>
> Using write-behind translator should improve small writes performance,
> but using it in configuration with dht over two afrs shows to me almost
> none improvement. When I'm watching networkt traffic (gluster works on
> 1Gb network) for every send packet there is one recieved:
>
> 10:05:20.205346 00:1e:c9:b9:bc:75 (oui Unknown) > 00:1e:c9:b9:bc:39 (oui
> Unknown), ethertype IPv4 (0x0800), length 189: n1.1019 > n2.6996: P
> 138505:138628(123) ack 71706 win 6533 <nop,nop,timestamp 517009620
> 517010547>
> 10:05:20.205583 00:1e:c9:b9:bc:39 (oui Unknown) > 00:1e:c9:b9:bc:75 (oui
> Unknown), ethertype IPv4 (0x0800), length 170: n2.6996 > n1.1019: P
> 71706:71810(104) ack 138628 win 3027 <nop,nop,timestamp 517010547 517009620>
> 10:05:20.205746 00:1e:c9:b9:bc:75 (oui Unknown) > 00:1e:c9:b9:bc:39 (oui
> Unknown), ethertype IPv4 (0x0800), length 126: n1.1019 > n2.6996: P
> 138628:138688(60) ack 71810 win 6533 <nop,nop,timestamp 517009620 517010547>
> 10:05:20.205801 00:1e:c9:b9:bc:39 (oui Unknown) > 00:1e:c9:b9:bc:75 (oui
> Unknown), ethertype IPv4 (0x0800), length 110: n2.6996 > n1.1019: P
> 71810:71854(44) ack 138688 win 3027 <nop,nop,timestamp 517010547 517009620>
> 10:05:20.205945 00:1e:c9:b9:bc:75 (oui Unknown) > 00:1e:c9:b9:bc:39 (oui
> Unknown), ethertype IPv4 (0x0800), length 158: n1.1019 > n2.6996: P
> 138688:138780(92) ack 71854 win 6533 <nop,nop,timestamp 517009620 517010547>
> 10:05:20.205994 00:1e:c9:b9:bc:39 (oui Unknown) > 00:1e:c9:b9:bc:75 (oui
> Unknown), ethertype IPv4 (0x0800), length 110: n2.6996 > n1.1019: P
> 71854:71898(44) ack 138780 win 3027 <nop,nop,timestamp 517010547 517009620>
> 10:05:20.206145 00:1e:c9:b9:bc:75 (oui Unknown) > 00:1e:c9:b9:bc:39 (oui
> Unknown), ethertype IPv4 (0x0800), length 126: n1.1019 > n2.6996: P
> 138780:138840(60) ack 71898 win 6533 <nop,nop,timestamp 517009620 517010547>
> 10:05:20.206211 00:1e:c9:b9:bc:39 (oui Unknown) > 00:1e:c9:b9:bc:75 (oui
> Unknown), ethertype IPv4 (0x0800), length 110: n2.6996 > n1.1019: P
> 71898:71942(44) ack 138840 win 3027 <nop,nop,timestamp 517010547 517009620>
> 10:05:20.207543 00:1e:c9:b9:bc:75 (oui Unknown) > 00:1e:c9:b9:bc:39 (oui
> Unknown), ethertype IPv4 (0x0800), length 168: n1.1020 > n2.6996: P
> 115271:115373(102) ack 61422 win 6161 <nop,nop,timestamp 517009620
> 517010525>
> 10:05:20.208136 00:1e:c9:b9:bc:39 (oui Unknown) > 00:1e:c9:b9:bc:75 (oui
> Unknown), ethertype IPv4 (0x0800), length 245: n2.6996 > n1.1020: P
> 61422:61601(179) ack 115373 win 19825 <nop,nop,timestamp 517010547
> 517009620>
> 10:05:20.208291 00:1e:c9:b9:bc:75 (oui Unknown) > 00:1e:c9:b9:bc:39 (oui
> Unknown), ethertype IPv4 (0x0800), length 66: n1.1020 > n2.6996: . ack 61601
> win 6161 <nop,nop,timestamp 517009620 517010547>
>
> In config there is:
>
> volume wb
>        type performance/write-behind
>        option block-size 256kB
>        option cache-size 2MB
>        option flush-behind on
>        subvolumes iot
> end-volume
>
> So shouldn't write-behind translator agregate these small writes in one
> bigger - small packets in larger or larger window?
>
> Regards
>
> --
> rash at konto pl
>
> _______________________________________________
> Gluster-users mailing list
> Gluster-users at gluster.org
> http://zresearch.com/cgi-bin/mailman/listinfo/gluster-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://supercolony.gluster.org/pipermail/gluster-users/attachments/20090206/1890b3e0/attachment.html>


More information about the Gluster-users mailing list