Jim Rosinski (rosinski@bearmtn-e0.cgd.ucar.edu)
Mon, 25 Jan 1999 10:16:44 -0700 (MST)
On Sun, 24 Jan 1999, Frank Peacock wrote:
> I am intending to run CCM3 on a PC-based machine. I am not sure of the
> relative importance of cache size to processor speed for improved
> forecasting time.
>
> The present considered machines are either:
>
> a 450MHz Xion with 512Kb cache OR
>
> a 400MHz Xion with 1 Mb cache.
>
> Is there a simple answer to this dilemma or has anyone tested something
> along these lines?
I don't have a complete answer to either question. But isn't the 400MHz
vs. 450MHz processor cost differential around $250, and the 512Kb vs. 1 Mb
cache cost differential around $4,000? Unfortunately I cannot promise that
paying the extra $$ for the cache will be worth it. Main memory bandwidth
has proven to be the primary limiting factor affecting performance for most
architectures on which CCM3 runs.
> Since our memory setup will be totally in-core does this permit the cache
> to be utilized and is the spectral data ordering scheme for non-vectored
> computers possible, with no added programming, on a cached PC-machine?
The answer to both questions is yes. The CCM3 code as run on non-vector
machines is "cache-friendly" in that most of the inner loops (including
spectral dynamics) are stride 1. However, the code is "cache-unfriendly" in
that most inner loop lengths are long (that most inner loop lengths are long (128@T42) and contain lots of work,
while less than abundant opportunity exists for cache reuse.
I am very curious to hear the outcome of what you decide to buy, and how well
the code performs. I can provide timing benchmarks from other machines if
and when the time comes.
Regards,
Jim Rosinski
CCM Core Group
This archive was generated by hypermail 2.0b3 on Tue Aug 24 1999 - 11:27:10 MDT