[Tfug] CPU Query

Jim Secan jim at nwra.com
Sat Mar 31 16:06:21 MST 2007


Robert Hunter wrote:
> You would almost certainly get a bigger gain by
> introducing parallelism into your code. E.g., with threads,
> auto-parallelizing compilers, libraries such as scalapack, etc.  I can
> understand why you would be reluctant to make such an investment of time
> to retool, but in the long term it may very well pay off.  It looks like
> the trend toward multicore processors is manifest.  Already Intel is
> marketing quad cores, and AMD will soon follow.  I'm certain I will be
> picking up some of AMD quads as soon as they come out.

I don't doubt that this could be optimized to run faster on multiple
processors by code parallelization.  Thing is, there's tradeoffs.  I know
the code as it is now is correct and doesn't produce weird results (well,
most of the time).  If I retool it to optimize for running on multiple
processors, there's a code/test/swear/recode/retest loop of unknown
dimensions before I'm back to where I am now.  I don't have the time to do
this now.  I also know there are compilers that tout themselves as
hands-off optimizers, but 35+ years in the business makes me wary of such
claims.

If I can buy a 2X, or even buy two 2X boxes and split the processing
between the two (I'm processing many sets of data, and each data set is
independent of the others), I can for very little time investment run
things through X times faster than I could before.  That's why I'm
interested in what a single 2X will give me running the same code on a
cheaper 1X of the same clock speed.  I don't think I'll get twice the
throughput, and I'm hoping it would not be less throughput, but I'd like
to know where between no difference and factor-of-two the truth is.  In
the "good old days" when my company was trying to decide between Sun (4),
SGI, Apollo, IBM, and a handful of other workstation-class computers, we
just trotted over to the vendors office with a tape full of codes and ran
benchmarks on the very codes we would be using.  Can't do that as easily
(if at all) these days.

And thanks for the cache suggestion.  I'll watch for that.

Jim





More information about the tfug mailing list