ScaleMP on Dell Servers!

I’m not sure how many people have looked at ScaleMP but they provide a unique product that allows you to take distributed nodes, connect them with a network, and make them appear and function as a single SMP machine. Today, ScaleMP announced the availability of their vSMP Foundation Standalone software on the Dell PowerEdge server platforms. In particular, they announced that vSMP is available for Dell’s new M1000e blade chassis allowing you to combine up to 16 blades (32 sockets or 128 cores) and up to 1 TB of memory into a single SMP machine. They also announced that support for their entry-level solution that allows you to connect two Dell PowerEdge 1950III servers with the vSMP Foundation Standalone software to create a high-performance virtual four-socket system.

I’m not sure how many people have examine what vSMP does, but let’s take a look at the Dell blades in particular. vSMP now supports the M1000e that can have up 16 blades (32 sockets) and up to 1 TB of memory. Using vSMP, when you boot the nodes in the M1000e, they boot up and the entire set of nodes looks like a single machine with 32 sockets! So it’s a single large SMP machine with a single instance of the OS. This SMP machine can use all of the memory controllers on all of the nodes, all of the peripherals on the nodes such as the disk controllers and the NICs, like it was a single machine.

This also means that you can allocate memory that is larger than the amount of memory on a single node! Many people have been asking for x86 SMP systems with large amounts of memory and with vSMP and Dell you can now have up to 128 cores and 1TB of memory in a single x86 SMP machine. If you write threaded code, you can now have up to 128 threads and they will be allocate to their own core!

Why have people been asking for large memory machines? There are several reasons. In some cases, there are applications that don’t scale well on distributed systems (clusters). But they do very well on SMP machines. Dell blades and vSMP allows you to get up to 1TB of memory on an x86 SMP machine. In other cases, some applications just need lots of memory even of the code is serial. For example, some grid generators for CFD codes are serial applications, but for large grids they need lots of memory larger than what you can get in a single 4-socket or 8-socket x86 system today.

Currently, vSMP works with Linux. But the requirements for the blade system hardware are the same as for a distributed system and are fairly simple. You need the chassis and the blades using however many blades you want (up to 16) with memory, whatever disks you want in the nodes. The blades also need to have an InfiniBand HCA in them. And finally, the chassis needs to have the Dell embedded InfiniBand switches.

For the PowerEdge 1950III solution, you just need two 1950III nodes configured however you like, an InfiniBand card in each node, and a short InfiniBand cable connecting the two nodes.

Jeff