Parallel MCR Performance Degradation

Is there any resource sharing among MCRs, such that when a large number (>> 7) of parallel instances (processes) runs on the same machine (Windows 7 x64, 24 cores, 20G RAM and Matlab 2011b x86) it degrades the MCR performance? We are experiencing huge performance degradation in our system as we go beyond the 7th instance (isolated processes) and would like to understand why that is happening.

Run-time data from our experiment (Processes - Run-time in seconds)

• 1 - 51
• 4 - 55
• 8 - 83
• 12 - 135
• 16 - 115
• 20 - 160
• 24 - 222

We would expect to see a small linear like degradation as we increase the number of parallel instances or processes, however we’re experiencing an unacceptable run-time loss.

Thanks for your help in advance.

Answer 1

[Jason Ross]

Cores aren't the only resource that can be exhausted. You could be using all the RAM, waiting on disk I/O, out of network bandwidth, etc.

You'll need to use a tool like the Resource Monitor (you can get to it from the Task Manager) to look at the other subsystems to see if they are maxed out.

[Israel]

Thanks for your input, I did check the Resource Monitor to try and identify the problem, but the hardware resources seems fine:

1 - each process requires about 300MB of RAM, we have 24GB available;

2 - there is no network bandwidth requirement, all running locally;

3 - disk I/O looks minimum with almost no activity.

As soon as we go beyond 7 or 8 parallel processes, the times deteriorate. The processes are computational intensive but there is no communication involved during run-time, apart from the simple process manager.

[Jason Ross]

Great. Also, are these 24 compute cores or 12 compute cores? Given how hyper-threading works, I've not seen benefit in exceeding the number of compute cores if you are doing computationally expensive work.

I'd expect that degradation to start at the 13th instance, though.