SMP and Network Connections

[Joseph Mack]

Douglas Eadline wrote:
> 
> Just posted some more SMP tests on www.cluster-rant.com.
> This time, I tested the interconnects and asked the
> question "What if a dual SMP used two Ethernet connections
> instead of one?" Seems to help! Take a look at:

Thanks for your work and write up. 

I did some performance tests a few years ago on a router,
using multiple copies of netpipe through a single interface 
to a set of nodes, to determine the effect of multiple streams
on throughput on the router (this was 100Mbps ethernet).

I found that as I increased the number of nodes connecting
to the router, that the throughput increased, rising above
100Mbps (when I totalled the throughput from each netpipe
job).

Looking at the netpipe code I saw that netpipe waits for a
quiet time on the network before entering the next round of
the test. Thus for 4 connections, if each instance of
netpipe waited for a quiet time to run the test on the next
packet size, I could (in principle) get the result of 4 connections
of 100Mpbs for a total of 400Mpbs. 

I contacted the netpipe author, who sent me a preliminary version
of a multi-netpipe, where multiple connections are synchronised
and stepped through the range of packet sizes together. He
said that it wasn't ready to use and I didn't have time to 
work on it myself. I never solved the multiple connection 
problem and wound up doing tests with a single connection.

Do you know if this problem is affecting your measurements?

(The report is at
http://www.linuxvirtualserver.org/Joseph.Mack/performance/single_realserver_performance.html)

Joe

-- 
Joseph Mack PhD, Senior Systems Engineer, SAIC contractor 
to the National Environmental Supercomputer Center, 
ph# 919-541-0007, RTP, NC, USA. mailto:mack.joseph at epa.gov
_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf