"Brian W. Barrett" wrote:
> On Wed, 4 Sep 2002, Jose Luis Bosque wrote:
>
> > I am working with a cluster with 8 Pc's, and I need to measure the
> > overhead time of the MPI_Recv function, in a point to point
> > communication, for messages of different sizes. With messages size under
> >
> > 60000 bytes there are no problems. But when the messages are longger,
> > the MPI_Recv functions as a synchronous function, and the values of
> > overhead time are not correct. Is there any way to change this limit ?
>
> Everything that Roberto said is true - you can change the cross-over point
> at configure time. This will push back the point at which you see the
> change from "fire and forget" short messages to "request to send" longer
> messages.
>
> However, I question your logic in doing this (and therefore the value of
> the benchmark data you will produce). There are performance changes that
> result from moving the short/long switch-over point. 64k is the default
> because it generally matches the maximum buffer size for a TCP connection.
> This is one of those points where bigger is not necessarily better for a
> real application. LAM makes some assumptions about sending short messages
> that definitely would not be true for very large message sizes. This
> would cause "much badness" for many applications.
>
> So, setting the cross-over point to a very big number for normal MPI use
> is obviously not a "good thing". So you don't want to do this in a
> benchmark suite - benchmarks that flagrantly bend normal operating
> conditions are even less useful than normal benchmarks :).
>
> I'm not really sure what you mean by overhead - do you mean the entire
> time that the benchmark app is in the MPI function but not actually
> sending data? Or do you mean the entire time that the benchmark app is in
> the MPI function? Or something entirely different? I'm sure that we can
> come up with a way for you to get your numbers, but I'm not sure what you
> need. Let me know, and I'll help you come up with something that will
> more accurately reflect the operating conditions of LAM/MPI.
>
> Brian
>
> --
> Brian Barrett
> LAM/MPI developer and all around nice guy
> Have a LAM/MPI day: http://www.lam-mpi.org/
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
Hi Brian:
Thanks by the reply and your interest, and also thanks to Roberto. I am
developing a parallel computational model for heterogeneous clusters, based
on LogP model. I need verify the accuracy of the model, so I need measure all
the parameters (Latency, overhead and gap) for a cluster. The parameter "o"
is the overhead of the communication function (so there are an overhead for
send and a different overhead for receive). This parameter is defined in the
model as "the lenght of time that a processor is engaged in the transmission
or recepcion of a message".
As I said in the last mail, I can measure the overhead of send
function. But the problem is with the receive function and only with long
messages.
Thanks again.
Jose Luis.
_______________________________________________
This list is archived at http://www.lam-mpi.org/MailArchives/lam/
|
