[Beowulf] split traffic to two interfaces for two "subnets"

Dan Stromberg strombrg at dcs.nac.uci.edu
Thu May 11 14:53:42 PDT 2006


If you have an application that's sending data into the IP stack in
chunks < 9000, then the 9000 MTU may not help that much, barring
something like Nagle.

However, if you crank up your NFS wsize and/or rsize, you may see a nice
boost.

NFS isn't known as a stellar performer in terms of network performance.

On Thu, 2006-05-11 at 17:12 -0400, Yaroslav Halchenko wrote:
> Just thought that it might be interesting for someone
> 
> ok -- I setup bonding ;-) and it works ;)
> with my tuned net/ip params:
> 
> net.core.wmem_max = 524288
> net.core.rmem_max = 524288
> net.ipv4.tcp_rmem = 4096 524288 524288
> net.ipv4.tcp_wmem = 4096 65536 524288
> 
> and with mtu 9000 on both ends, running two netperfs (netperf params
> are  -P 0 -c -C -n 2 -f K -l 20 ) to two nodes give me next
> results:
> 
> (in KBytes/sec... why didn't I make it in megabits? :))
>          node1           node2           total
> average	64276.88	82695.07	146971.95
> std	20555.99	20215.57	10685.59
> min	29857.41	39972.52	129420.64
> max	110149.35	112383.19	166813.32
> 
> that is simplex -- ie just netperf to the nodes...
> so, the results are not that great in comparison to 2 split interfaces
> where each one can achieve up to max you observed in the table
> 
> may be I should have tried different MTUs... at least for duplex traffic
> MTU around 3500 provide better performance...
> 
> since I am aiming at NFS I would also better check bonding performance
> on NFS tests...
> 
> after all it is  a cheap router ;-)
> 
> On Thu, 11 May 2006, Yaroslav Halchenko wrote:
> 
> > Great great great news for me ;-)
> 
> > Thank you Joshua -- I will try to set things up and use trunking
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf




More information about the Beowulf mailing list