All our SM 'twins' are 8-core beasties (2 x 4-cores) per side. They all
sit on a 1GbE network (Nortel 5510/5530 etc).
To answer you question - yes there will be contention if one or more
jobs want to do i/o to the net at the same time. This happens whatever
the speed of the NIC. If you have a 100Mb/s NIC, then in general,
everything will take ~10 times as long than for a 1Gb/s NIC for a single
process. If you have two processes, the transaction(s) will take >10
times as long because of context switching between processes for the
resource (the NIC).
If your jobs are running several tens of hours, then a 100 second data
squirt at the end is insignificant compared to the run time - and making
it 200 seconds is still insignificant. However for high-i/o jobs where
the i/o is 'continuous' then it could double your run time.
It may be that your data sink is the limiting factor - if it is a 1GbE
system, then you could get 10 x 100Mb/s data streams off it (assuming
that it can get the data on/of the disk that fast).
The best strategy is to force the jobs to do their I/O to the local disk
and and then squirt it off in one lump at the end of the job. The input
data has to be pre-staged locally too. Then the normal smearing out of
start and finish of jobs due to slightly different run times spreads the
high load to the servers as well.
Martin.
--
Martin Bly
RAL Tier1 Fabric Team
> -----Original Message-----
> From: Testbed Support for GridPP member institutes
> [mailto:[log in to unmask]] On Behalf Of Winnie Lacesso
> Sent: 29 August 2008 13:59
> To: [log in to unmask]
> Subject: 8-core WN with single NIC - contention?
>
> Question for those of you with 8-core WN, where something like
> gpfs is *not* in the equation - do you have these on a 1Gbps
> network or a 100Mbps network?
>
> Because one wonders, if WN are on a 100Mbps network could there be
> contention between several high-I/O jobs on that WN for that single
> 100Mbps eth0 NIC?
>
> Anyone seen this?
>
|