Dear all,
We are running Matlab 2011a on a shared Windows 2008 Terminal Server (TS for short). The users login to the TS from their workstations using Microsoft Remote Desktop. The TS is a physical box (8 Core(s) x 4 2260MHz), 250GB memory.
Typically we have 5-6 users of Matlab active on the system at any one time. The users edit and debug their Matlab code through the Matlab GUI. For heavy calculations we provide a cluster of Blade servers running MDCS - so the users can offload and get decent performance.
On an apparently sporadic basis the Matlab GUI users report that the Matlab GUI becomes unresponsive - i.e. clicking on menus and typing in the command-window is slow. I have dome some performance tracing on the system at the time the users report the problem - my empirical observations have been: - Overall system CPU and memory usage at the time the problem occurs are within "reasonable" bounds - i.e. no more than 25% - There will be 4 or 5 MATLAB.EXE processes running (one per user) - and maybe one of them will be consuming 20GB memory because the one user is doing some kind of big array manipulation. On average the others are consuming a few hundred MB each. CPU consumption no more than 2-3%. - Other aspects of the remote desktop (e.g. Windows Start Menu, Notepad, MS Office Applications, SAS Enterprise Guide, Eclipse etc) behave quite responsively - and do not seem to be impacted even though Matlab is frozen.
As a test, I fired up several 5 MATLAB GUI sessions in parallel, and in each one launched a calculation that is manipulating a large array (80MB) and in a 6th MATLAB GUI session tested the responsiveness of the GUI. Sure enough performance dies - even though on the system overall, CPU and memory are hardly used.
My current take on the situation:
My understanding of the MATLAB GUI is the software (MATLAB.exe) is little more than a Matlab MDCS worker running in GUI mode - or conversely a MDCS worker is MATLAB.EXE running in non-GUI mode.
When we first built the MDCS compute cluster we did it on a 32 core W2K8 application server - and found that once you got beyond 6 cores being occupied by (e.g.) a parfor loop that is making intensive manipulation of arrays, matrices etc, we ran into a bottleneck caused by contention on L3 cache. Hence we got around this by basing our cluster on many smaller 12-core Blade servers. and we run no more than 6 workers per server.
I am wondering if we are seeing a similar L3 contention issue on the Terminal Server - and maybe the answer is to move to several smaller terminal servers. But the fact that other applications peform fine when Matlab is the one hanging / going slowly makes be doubt this.
We are already in conversation with Mathworks to discuss the problem. Although the installation notes for Matlab state that Windows 2008 is supported - Mathworks support take the line that they have never really tested Matlab GUI running on a shared Terminal Server. To be fair they are helping us to look into the problem nevertheless.
My questions to the forum are:
- Is anyone else out there running Matlab on a shared Terminal Server similar to what we are doing?
- If yes, have you encountered similar performance issues to what I describe where Matlab just hangs from time to time?
- Did you manage to find a workaround?
Thanks in advance.
Cheers,
Nick Steven.
Zürich, Switzerland
