[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: mpi performance consideration

At 12:02 PM 11/16/98 +0100, you wrote:
>Hello all,
>We have a cluster of 16 LX164 533Mhz on a 100 Mb fd switch.
>We have bad performance with mpich 1.0.9 or lam 6.1 with both 2.0.35 or
>kernel with eth drivers de4x5.c:V0.542 100Mb/s full duplex or
tulip.c:v0.83 10/19/97
>I include the small program latband.c, tested on two nodes, compiled with
>gcc version egcs-2.92.17 19981026 and glibc 2.0.7
>with 2.0.35 lam6.1 -c2c, we have around 64 µsec.
>without -c2c, it's 300 µsec or worse, with all the combinaisons of mpi distr,
> kernel, eth ...
>If someone have better performance or any idea in order to improve it,
>please feel free to contact myself.
>(With OSF 4.0, it's around 30 µsec so the users want to switch to OSF :-| )
>Bernard Van Rent 

Well, I'm not an expert on MPI/clustering, but I'll throw in one
suggestion. On almost all Beowulf clusters that I have seen, the inter-node
bandwidth is usually 200-400 mbps. Is the limitation of your system the
inter-node bandwidth? Have you thought of adding bonded ethernet, and
perhaps another switch? If it's not a limitation there, I'd love to hear an
explaination from someone as to what exactly is happening.



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index] []