Newsgroups: comp.parallel.mpi
From: raja@convex.com (Raja Daoud)
Subject: Re: MPI-Benchmarking
Organization: Hewlett-Packard Co.; Convex Division
Date: 27 Oct 1996 20:42:40 -0600
Message-ID: <5516j1$63h@tbag.rsn.hp.com>

Peter Sanders <sanders@ira.uka.de> wrote:
>Why we think this is useful:
>- Programmers would like to be able to estimate the consequences of
>  implementations decisions before they actually run the code, because [...]

[...]

>- [Will not] Collapse the entire results into a single "performance index"
>  because this might be misleading.

With these two goals in mind, I'd like to request the following:

>- prepare and evaluate the data in tabular, graphical and
>  mathematical form (by fitting curves)

When doing the curve fitting for point-to-point functions, treat
piece-wise linear behaviour as such and don't force (collapse) the
data to fit the single line model.  Likewise for collective routines,
separate the linear and logarithmic sections of the curve where
applicable.

>- Are there any clever tricks for avoiding mismeasurements due to
>  operating system intervention, cache effects,...(beyond taking
>  averages/medians etc.)

Some OSes have lazy memory allocation.  To factor this out, do a
"warm-up" exchange or initialize the data buffers before timing.
Otherwise, the first exchange takes a long time because it includes
the memory allocation time.


--Raja

-=-
Raja Daoud					Hewlett-Packard Co.
raja@rsn.hp.com					Convex Division

