[Beowulf] SCI Socket latency: 2.27 microseconds

Bill Broadley bill at cse.ucdavis.edu
Mon Mar 1 16:38:52 EST 2004

> well, Quadrics now claims 1.8 us MPI latency:
> http://doc.quadrics.com/quadrics/QuadricsHome.nsf/NewsByDate/5D7D7444A42B398280256E44005A31DD

Note the title says "sub 2us" and the body says "close to" 1.8us.

Of more interest (to me) is that further down they say:
 In the next quarter, Quadrics will announce a series of highly competitive
 switch configurations making QsNetII more cost-effective for medium
 sized cluster configuration deployment.

Sounds like more competition for IB, Myrinet and Dolphin.  Hopefully anyways.

Cool, found a quadrics price list:

Looks like $3k per node or so for 64, and $4k per node for 1024, I'm guessing
that is list price and is somewhat negotiable.

According to my sc2003 notes the Quadrics latency was:
	 100ns for the sending elan4
	 300ns for the 128 node switch and 20 meters of cable
	 130ns for the receiving card.
	2420ns for two trips across the PCI-X bus and a main memory write
    2950ns for an mpi message between 2 nodes.

Anyone know what changes to get this number down to 1.8us - 2.0us?

> higher bandwidth.  yes, some vendors (esp IB) are pushing PCI-express
> as bandwith salvation, but afaikt, none of my users need even >500 MB/s
> today.  it doesn't seem like PCI-express will be any kind of major win
> in small-packet latency...

Anyone have an expected timetable for PCI-express connected interconnect

Anyone have projected PCI-express latencies vs PCI-X (133 MHz/64 bit)?

Bill Broadley
Computational Science and Engineering
UC Davis
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