[Tfug] Processing/memory throughput and suggestions

[Zack Williams]

> But, it occurs to me that I might possibly achieve better throughput if I went
> with a true dual-processor motherboard and put onto it two dual-core CPUs.
> More paths to the memory means less contention inside the CPUs, but, of
> course, there can still only be one thing accessing memory at a time.

This totally depends on memory architecture of the system. The
suggested Core2 system would not be a good choice for this, as memory
access for the processors is bottlenecked in the chipset.

For all recent AMD systems, or the Intel Core i7, the memory is
attached directly to the CPU - adding more sockets increases the
number of memory channels, and thus your memory bandwidth for the
entire system.  It also helps if the OS has proper NUMA support and
can assign a job to one CPU/memory combination.

> Anybody have experience with this sort of thing?  The jobs the machines will
> run are quite varied.  Some are processor bound and run for a long time.
> Others are memory bound, using huge amounts of memory (large matrices) and
> running over and over the memory (such that the cache isn't as useful as it
> could be).

I'd look into a Core i7, or the Xeon 35xx or 55xx series system, or a
quad core AMD Phenom II or Opteron x3xx series processor.

With that quantity of RAM and job runtimes, you'll want to look into
ECC options to prevent/detect memory errors.

> I'd also welcome any suggestions for motherboard/CPU/memory combinations that
> you think might be better.

What price range are you aiming for?

For the upgrade potential, I'd look at a dual socket workstation
board, but buy only one processor and set of memory now and the other
at a later date.

- Zack