Gigabit Ethernet switches and network adaptors.

Velocet math at velocet.ca
Tue Nov 20 13:25:51 PST 2001


On Tue, Nov 20, 2001 at 02:55:07PM -0600, W Bauske's all...
> Steve Gaudet wrote:
> > 
> > Hello,
> > 
> > > On Mon, Nov 19, 2001 at 10:00:46AM -0500, Steve Gaudet's all...
> > > > Here's some of the network GIG E hardware we'd like to recommend:
> > > >
> > > >   AceNIC/NetGear GA620(T)/3C985B
> > > >   SysKonnect
> > > >   NS chipset:
> > > >      Cameo           SOHO-GA2000T    SOHO-GA2500T
> > > >      D-Link          DGE-500T
> > > >      PureData        PDP8023Z-TG
> > > >      SMC             SMC9462TX
> > > >      NetGear         GA622
> > >
> > >
> > > How do you find the performance of these NS82830 cards? Do they do
> > > block interupt xfer or whatever it is for more efficient
> > > xfer? How much
> > > system/interupt time do they chew up?
> > 
> > The NS based stuff is low-end and cheap - you get what you pay for - but the
> > drivers are rock solid and provide a good way to get started with GigE.
> > 
> > On a high enough powered box, you might even get decent throughput but it's
> > at a
> > cost of cycles.
> > 
> 
> If you consider 57MB/sec for a $45 card bad, then the ns83820's
> are a bad deal. I consider that a good buy. CPU is around
> 30% of a P4 1.5ghz system, both sending and receiving sides.

I dont know what CPU was in use at the time, but I've gotten over 350Mbps
out of them with

one machine: 

dd if=/dev/zero bs=1M count=1000 | nc -w 1 othermachine 33333

othermachine:

nc -w 1 -l -p 33333 | dd of=/dev/null 

(dd on freebsd tells Bps on stderr on termination or interupt) 

> It would be nice if the cpu was lower but it's acceptable for what
> I do. (YMMV) 

With gromacs going between 2 dual-athlon 1.33Ghz CPUs running the d.dppc 
benchmark I only noticed about 2-3% system time and 87% usertime. (Linux 
doesnt seperate interupt and system time like freebsd does, and I
was using linux in this test).

> Also would be nice if Gbe switches were cheaper. Latest pricing I 
> see is $600 for an 8 port switch, or $75/port+$45/card for $120
> per connection. 

Why not just go for SCALI? Since when does COST matter on the Beowulf 
list? </snide> :) 

/kc
--
Ken Chase, math at velocet.ca * Velocet Communications Inc. * Toronto, CANADA 



More information about the Beowulf mailing list