[SCore-users] Help: Too slow running any program
Bogdan Costescu
bogdan.costescu at iwr.uni-heidelberg.de
Thu Jul 31 21:07:54 JST 2003
On Thu, 31 Jul 2003, Huiqun Zhou wrote:
> It seems that programs running in my Score environment not only took
> long, long time between "Score-D connected" and the first line of output
> displayed, but also spent more time on communication during computation.
Then most likely the network is not configured properly, some packets are
lost and because of the retries the run time increases very much. I think
that you need to check the network: start by using netperf or ttcp or
something similar which can give you an idea about capabilities of the
network in terms of (TCP or UDP) transfer speed. Check the duplex
settings of the switch and of the network cards: for the cards, a tool
like mii-tool (comes with most distributions, including Red Hat) or
mii-diag (ftp://ftp.scyld.com/pub/diag) will tell you if the media was
negotiated or not and what are the current values; if the switch is
manageable, the control interface will give the same data - if not, you'll
have to assume that it does what it should, which normally means that it
tries to autonegotiate. Also check cables, I've seen puzzling problems
because of bad cables. If you are using the 3c59x driver, you can ask for
more help by writting to vortex at scyld.com.
> it took 4 to 6 seconds per time step even on a 7 compute node cluster!
Please compare apples to apples! If your previous results come from a
4-node cluster, please give new results also from a 4-nodes cluster (or a
run using 4 nodes on a larger cluster). Otherwise the differences might
very well be explained by the Amdahl's law.
--
Bogdan Costescu
IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu at IWR.Uni-Heidelberg.De
More information about the SCore-users
mailing list