[Beowulf] SCI Socket latency: 2.27 microseconds

Friedrich Seifert sfr at foobar-cpa.de
Tue Mar 2 04:40:46 EST 2004

Bogdan Costescu wrote:

> On Mon, 1 Mar 2004, Petter Gustad wrote:
>>Dolphin has benchmarked a completed one byte socket send/socket
>>receive latency at 2.27 microseconds,
> Is this in polling mode or interrupt-driven ? I'm interested to see if
> I can do something useful (like computation) _and_ get such low
> latency.

Actually, SCI SOCKET uses a combination of both, it polls for a 
configurable amount of time, and if nothing arrives meanwhile, waits for 
an interrupt. Something like that is necessary since the current Linux 
interrupt processing and wake up mechanism is quite slow and 
unpredictable. There is a promising project going on to provide real 
time interrupt capability, but it is still in an early stage 

>>Benchmarks using Netperf also show more than 255 MBytes (2,035
>>Megabits/s) sustained throughput using standard TCP STREAM sockets.
> What is the CPU usage for this throughput ?

SCI SOCKET was run in PIO mode for this test, so one CPU is needed to 
transfer the data. Current DMA performance is lower, but is subject to 
optimization in future revisions. CPU usage for DMA is 8%/29% at 


Dipl.-Inf. Friedrich Seifert - foobar GmbH
Phone: +49-371-5221-157         Email: sfr at foobar-cpa.de
Mobil: +49-172-3740089          Web:   http://www.foobar-cpa.de

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