From: nmm1@cus.cam.ac.uk (Nick Maclaren)
Newsgroups: comp.parallel.pvm,comp.parallel.mpi
Subject: Re: Switching HUB for beowulf cluster?
Date: 4 Feb 1999 12:15:57 GMT
Organization: University of Cambridge, England
Message-Id: <79c31t$jm6$1@pegasus.csx.cam.ac.uk>
References: <78osd7$9ji$1@news.kren.nm.kr>
    <78piue$lrv$1@pegasus.csx.cam.ac.uk> <36B966E6.F5E27A60@dolphinics.com>
    <79c0na$hk5$1@pegasus.csx.cam.ac.uk> <79c1vd$5lm$1@news.uni-paderborn.de>
Originator: nmm1@taurus.cus.cam.ac.uk
Xref: ukc comp.parallel.pvm:7970 comp.parallel.mpi:4587


In article <79c1vd$5lm$1@news.uni-paderborn.de>, butenuth@lilienthal.uni-paderborn.de (Roger Butenuth) writes:
|> In article <79c0na$hk5$1@pegasus.csx.cam.ac.uk>, nmm1@cus.cam.ac.uk (Nick Maclaren) writes:
|> |> 
|> |> The current cost-effective solution is a 100baseT switching hub.
|> |> For better bandwidth (but not latency) at a higher cost, Myrinet,
|> |> HiPPI, Giganet etc. are useful.  For peak bandwidth, GSN (aka
|> |> HiPPI 6400) is the only game in town.
|> 
|> You can achieve high bandwidth and low latency with SCI (Scalable
|> Coherent Interface), we have measured about 81 Mbyte/s throughput and
|> 3 microseconds latency (in shared memory mode). Scali's MPI has about
|> 15 microseconds latency.

Hence my "etc."!  That performance is about the same as HiPPI and,
I believe, Myrinet.  GSN has been measured at several hundreds of
MB/sec.  3 microseconds DMA latency is fairly typical, whatever
networking technology you use.


Regards,
Nick Maclaren,
University of Cambridge Computing Service,
New Museums Site, Pembroke Street, Cambridge CB2 3QG, England.
Email:  nmm1@cam.ac.uk
Tel.:  +44 1223 334761    Fax:  +44 1223 334679