LAM and channel bonding doesn't work?
xyzzy at speakeasy.org
Fri Jul 13 15:15:07 PDT 2001
On Thu, 12 Jul 2001, Martin Siegert wrote:
> If there is somebody on the list who uses LAM and channel bonding: would
> you be willing to verify or disprove this? - You probably have to change
> the system size, I strongly suspect that the critical size is memory dependent:
> my numbers (4000 and 8000) are for 512MB of RAM.
Two dual CPU machines with 512MB RAM and two channel bonded fast ethernet
interfaces. Software is kernel 2.2.17, glibc2.0, lam 6.5.1-usysv.
[tap at venturi tap]$ mpirun -O n13,14 a.out -2 8000
id=0: MPI_Isend done.
id=0: MPI_Irecv done.
id=0: elapsed time: 12.901334 s
id=1: MPI_Isend done.
id=1: MPI_Irecv done.
id=1: elapsed time: 13.191415 s
Seems to work fine. Sending half of a 8000x8000 matrix of doubles in 15.65
seconds is about 18.9 MB/sec throughput in each direction. About what I get
for a netperf TCP throughput test.
More information about the Beowulf