Strange limits on Parallel RAM Usage by Compiled Standalone

I have a MATLAB-compiled standalone .exe on a desktop with 20 cores and 32GB RAM. Because of the ample resources of the desktop, I tried to launch several instances of the .exe in parallel. I find that any more than 2 instances running in parallel will crash with memory allocation error messages. However, each instance on its own consumes no more than 4 GB RAM according to the Windows Task Manager. With 32GB RAM total, I reason that 4 instances should easily be able to run side-by-side.

Any theories as to why I am limited to so few instances when I have so much RAM to spare?

[dpb]

Are these independent images or is there any shared memory/parallelism going on between processes?

[Matt J]

They read some common input files, but that's it. The outputs are completely separate.

[dpb]

Computations, though??? Any sort of problem during startup with contention for these common files? Are you checking that there's not a lock on the file when reading it in the app? What if you make copies for each to be sure--do symptoms change?

[Matt J]

I could try making copies, but it's hard to see why a locked file would lead to memory allocation errors... And the common file is just a .mat file with no unusual permissions set. If the file is being read by a particular process, I don't see why other processes wouldn't simply wait their turn for access to it.

[dpb]

I don't know, either -- simply wondering if might help uncover what's going on if symptoms were to change is all...

Answer 1

[Matt J]

A co-worker suggested a possible explanation. The MATLAB .exe is 64-bit compiled and is deployed on a 64-bit desktop. However, it is calling some external DLLs that were 32-bit compiled. The way I understand things from my co-worker, such DLLs execute in some virtual 32-bit Windows machine where the total available RAM is only 4GB. This 4GB RAM is shared by all other processes that happen to be running the DLL at the same time.

If the above is the true picture of how 32-bit DLLs run under 64-bit Windows, it would explain why memory ceiling is a lot more limited than I think it is. I welcome comments from anyone who can either confirm or refute this theory.

[Matt J]

As further evidence of the above, I am finding that I can run more than 3 parallel processes without difficulty, even when memory intensive, when I omit the steps that use the 32-bit DLLs. I'm closing the question based on this, but still welcome further theories.