Re: Trouble with 3c905 new cards (CX-TX)
Bogdan Costescu <[email protected]>
| Newsgroups | gmane.linux.drivers.vortex.devel |
|---|---|
| Message-ID | <Pine.LNX.4.44.0307101422390.3717-100000@kenzo.iwr.uni-heidelberg.de> |
On Wed, 9 Jul 2003, Simone Montagnani wrote: > I have a Linux redhat 7.3 Cluster (40 Biproc-nodes) with 3com 3C905CX-TX, I > have network performance problems Running the application was the only diagnostic that you have run ? You could try ttcp or netperf to see what the network is capable of, without any additional CPU or disk load. > When I run my application this 10 computers go really faster (I mean for > network wall time are equal to 40 nodes! ). But this is not a reason to suspect the network: parallel applications follow Amdahl's law and they don't scale linearly. Usually at some point the time taken to run on N CPUs is equal or even smaller than the time taken to run on 2*N CPUs; it means that you have hit the maximum number of CPUs that can be used for running this application in parallel (this is a very rough explanation, look up any parallel computing theory book for more details). > is this a driver problem?or a chipset card problem? I don't exclude the network card/driver as being the cause; maybe there are some problems with autonegotiation between cards and switch which lower the performance of network transfer when using the CX cards. But you have to find out if this is the case by getting more specific data about your network. -- 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: [email protected] _______________________________________________ vortex mailing list [email protected] http://www.scyld.com/mailman/listinfo/vortex