How to make all distributed nodes RAM available to a single node?

Question

I have an already developed application on Linux which would benefit from one process with more RAM than it would by off-loading individual processes onto smaller compute nodes with less RAM. So this is the one option I'm researching here.

I'd like to run one process on the master node and yet somehow make all the available RAM on all compute nodes appear as one large block of available RAM to the master node (as if a single system).

As far as I know, this use of a Single System Image (SSI) is now called Distributed Shared Memory (DSM). I've been looking into single system images but can't tell which ones support this.

Does anyone know which of the open source SSI packages would accomplish this? I was also wondering about OpenSSI, though the lists are very inactive.

Edit:
In my use of DSM, I was thinking in the context of having physically separate address spaces on different compute nodes, and being able to access them as if they were one logical shared address space available to a single process on a single node, the head node. I used the term SSI (and incorrectly in this context) because I was wanting the system to appear as one single system from the users perspective. My current use of SSI, is a file system in which all diskless nodes boot from a single image in a specific location on the master node.

I understand the memory of multiple nodes is accomplished by means of a comm lib (which is currently in use), but I need more RAM. So in researching SSI packages that seem to mention the ability to support a "Single IPC Space", I was thinking this might be something useful. I had read several white papers where the memory management of nodes could be handled at the OS level and therefore I wouldn't need the communication libary as this would be abstracted below user-space.

I'd like to not modify the code as it would already take advantage of large amounts of RAM if it were already available on the localhost. And I do have large amounts of RAM, except that RAM is distributed on different nodes and therefore does require the use of some MPI type libraries. Modifying the code is costly, adding more RAM requires a new cluster, costly. However, a solution that would give my application the illusion of a very large single address space ... I can only hope.

Answer 1

@Ender
You can download vSMP Foundation Free - it will allow you to aggregate the memory of up to 8 nodes, up to 1TB RAM total, free of charge. Note that the free version only allows for memory aggregation, meaning you will have the compute power of cores of only a single node, but the RAM of many.
http://www.ScaleMP.com/free

Answer 2

It's a commercial solution instead of open source, but I believe ScaleMP can do what you're looking for. I've worked at installations that used it in the past to aggregate multiple commodity servers into a larger system, but haven't used it directly myself. I think it requires Infiniband as a node interconnect, though, and they have a not-too-long list of supported hardware.

Answer 3

There are many choices that you can do including using MPI in a cluster, PGAS languages or using a software package like the one from ScaleMP. It really depends on your budget, how much your applications can take advantage of the different technologies and what your current application(s) are already written to use. If the apps run under a single OS only and don;t already take advantage of a clustered environment you best bet would be a SMP system (4 or even 8 CPU's) in a pure HW configuration or up to 128 CPU's in a virtual SMP configuration. Then you need to determine is it compute or just memory you need. If you go with the vSMP type implementation you can have 2, 4, 6 or more nodes that have high core count and high frequency CPU's. If you don't need the processing power, just the memory footprint you can add more nodes to the vSMP that have less expensive low core count and freq. CPU's and just use these as memmory controllers.

Answer 4

There are many libraries that accomplish this though you'd have to refactor your application to use them.

• Co-array Fortran
• Global Arrays
• OpenSHMEM
• Unified Parallel C

Just to name a few. None of these will be automatic, but it shouldn't be a huge burden to use them.