"Fast Switching" or NIC-to-NIC communications AND channel bonding discouragement

[Brian Dushaw]

On Mon, 13 Mar 2000, Bogdan Costescu wrote:
> 
> On Sat, 11 Mar 2000, Brian Dushaw wrote:
> 
> > On a similar topic, I have tried channel bonding with the 2.2.x kernels
> > (T. Davis' patch) and found that it was only marginally effective (and
> > not effective at all for MPI or pghpf parallelizations - things got worse).
> > There is some suggestion that this has to do with the "stack [??]"
> > inefficiencies of the 2.2.x kernels, that is better in 2.3.x [but what do
> > I know of this? nothing...]  Does anyone have any wise words for better
> > channel bonding performance?
> 
> How have you tested the channel bonding? Did you use netperf, ttcp or a
> similar tool?


I USED NETPERF.   ONE SET OF ETHERNET CARDS GAVE ME 140 MB/S, ANOTHER SET
GAVE ME 105 MB/S.  SINGLE CARDS GAVE ME 95 MB/S (SO IF CHANNEL BONDING
WERE 100% EFFICIENT ONE WOULD EXPECT 190 MB/S).  THE CHANNEL BONDING
SEEMED TO MAKE MY ETHERNET CONNECTIONS FAIRLY TEMPERMENTAL; I ABANDONED
IT.


> 
> You only get an improvement for MPI parallelizations if your software
> sends a lot of data, but not very often which means that you need more
> bandwidth and you don't care about latency. The channel bonding improves
> bandwidth allowing you to treat n NICs as one and thus increasing the
> bandwidth, but does not modify the TCP stack (which is responsible for the
> high latency). For more info about this subject, you should check
> www.beowulf.org and its associated mailing lists.


THANKS FOR THIS COMMENT.  IT APPEARS THAT CHANNEL BONDING IS NOT (YET)
APPROPRIATE FOR PARALLEL COMPUTATIONS, BUT IS O.K. FOR DATA SERVER.  THIS
IS WHAT MY EXPERIENCE WAS - MY MPI CODE TOOK TWICE AS LONG TO EXECUTE
(RATHER THAN HALF AS LONG...).  I CAN SEE THAT IT WOULD WORK FOR A DATA
SERVER.

> 
> I'm using channel bonding since its announcement on the beowulf list (that
> is, the new version 0.2 which is now included in the 2.2 series) without
> problems in a file-server.


I THINK THIS IS THE CODE THAT I USED AS WELL.

> 
> Best regards,
> 
> Bogdan Costescu
> 
> IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
> Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
> Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
> E-mail: Bogdan.Costescu@IWR.Uni-Heidelberg.De
> 

-------------------------------------------------------------------
To unsubscribe send a message body containing "unsubscribe"
to linux-tulip-request@beowulf.org