This comes on the heels of another announcement, a few days earlier, that the University of Florida’s Interdisciplinary Center for Biotechnology Research made the same decision. Both of these organizations will be using ScaleMP’s software to aggregate large amounts of memory that sit on distributed systems into a single, very large, memory space.
This will allow customers to run large jobs that just don’t parallelize that well. Being able to throw huge chunks of memory at their computing problems will aid performance, and being able to use the memory sitting in cheap(er) distributed systems should save some money to boot.
These virtual SMP systems can be brought up or torn down in 15 minutes or so. This means that customers can quickly and easily toss together big systems for particular jobs, rather than spend big bucks to buy a ready-made large SMP system – which would most likely be split up into smaller, virtualized chunks anyway. For customers with a need that only big iron can satisfy, this is the best of both worlds.
Using software wizardry along with fast Infiniband-based connections to make “on-demand SMP” big boxes is, well, pretty cool stuff indeed. It’s important to realize that this technology won’t work for every SMP workload – and, to their credit, ScaleMP is quick to point this out. They’re even happy to discuss where their product isn’t a good fit.
We had a quick briefing with ScaleMP last week and learned of some significant changes since we last talked with them earlier this year. They’ve seen interest in their vSMP Foundations product spread beyond the traditional research and university lab community and into the commercial space. They have notched a number of customer wins with folks in commercial markets who have found that they’ve reached the limits of cluster scalability and need large shared memory boxes to get the performance they need.
ScaleMP is seeing shorter sales cycles and deals which are mostly standard, meaning they don’t require a lot of handholding or implementation support. They also report that many of these commercial customers are using ScaleMP’s products as a foundation for their own private clouds.
To me, the news that ScaleMP is being picked up in the commercial world means that the virtual big iron concept is moving into a new phase. Several months ago, it was in the “Cool, let’s give it a try” phase (where universities and hardcore HPC-types will give it a tumble). Now it is being used in situations where the customers don’t have patience for solutions that don’t work out of the box – and work pretty much as advertised.
ScaleMP isn’t alone in this space, but it looks to have a lead at least in terms of mindshare and coverage. It will be interesting to watch this technology develop and mature. We could be seeing a real change in the way systems are designed and utilized. ®