ScaleMP

[dsidote]

We are considering whether ScapeMP's vSMP is comparable to a single box with 24 cores and 512GB of RAM. Does anyone have any first hand experience with ScaleMP?

Thanks.

[gardner]

Hi Dsidote,

Yes, we have a ScaleMP system we have been using for the past year or so for next-gen analysis.

As far as comparing to something like a Nehalem EX with 24 cores, 512GB of RAM I think the difference is in what you intend to use the system for. For running a couple similar jobs at a time with a large amount of shared memory and a low thread count, what you spec'd is a "safe" answer.

But when you "max out" a traditional SMP system like this, RAM is disproportionately expensive relative to the remainder of the system cost, processors are more expensive and slower, etc.

ScaleMP allows you to get flexible aggregate performance across cores, main memory, secondary storage, and HCAs. A comparable ScaleMP system to the one you spec'd would certainly provide you with more and faster cores and similar memory size and performance.

The thing to keep in mind is that informatics is a moving target, and what algorithms/programs are employed for one type of project change for the next, or change as instrumentation changes, or change because methods improve, or change for change sake.

Choosing a traditional SMP that is "maxed out" locks you in for however long you rely on that resource (usually years) with the thread count and memory you have. Alternatively, the ScaleMP system could be configured to have many more cores available from the start, and you can add memory and cores just by upgrading and adding more nodes (boards) and still have them all available in a single system image--whereas the "max" a single traditional SMP can provide is fixed for months at a time between hardware revisions.