Hardware requirements for MATLAB Distributed Computing Server 5.2 / mathworks job manager / parallel computing toolbox

Hi all

I have some queries about the hardware server requirements for MATLAB Distributed Computing Server 5.2 / mathworks job manager / parallel computing toolbox. I would appreciate very much if matlab technical support can help on them

1) For MDCS + mathwork job manager, they could be used together for distributing matlab tasks to computer cluster farm. If i want to rollout 10 x 8-cores workstations, is it a good choice? or should I use 3rd party scheduler like HPC? please advise.

2) the product specification of MDCS+mathwork job manager, states that hardware server running the workers requires “an NVIDIA Tesla Compute Cluster (TCC) driver with an NVIDIA Tesla card”. There are a few Tesla compatible workstations at(<http://www.nvidia.com/object/tesla_compatible_platforms.html>)

These workstations could have NVIDIA Tesla card (GPU graphic card) or general graphic card.

2.1) As you see, a Tesla card cost several thousand UDS. Is it a MUST to have the Tesla card installed at workstation so that matlab worker could be setup on the workstation, distributed task could be run? 2.2) If Tesla card is not a must and not installed, will it make a difference in term of matlab worker performance or matlab functionalities (E.g, some matlab programming instruction/function for GPU computing couldn’t work properly,right?)

3) Similarly, for parallel computing toolbox, if the desktop doesn't has CUDA-enabled NVIDIA GPU card, does it only affect GPU computing related functions? Can the tasks be distributed to local workers on multi-core desktop?

million thx for your answers. Danny

Answer 1

[Titus Edelhofer]

Hi Danny,

for the questions 2) and 3): the NVIDIA cards are only a requirement if you want to do GPU computing, i.e., use the GPU for doing some of the computations. If you stick to the "normal" workers running on the CPU (as MATLAB itself does), then this is not a requirement.

3) Yes, tasks can be distrubuted to local workers on multi-core desktops.

For question 1) I would contact MathWorks Sales who (in conjunction with the corresponding technical department) can handle those questions well.

Titus

Answer 2

[Jason Ross]

Just a note from personal experience -- if you are rolling out 10 machines to act as servers, then look into server hardware. Reason being is that they easily (and compactly) fit into a rack, have remote management features built in (power control, KVM, IMPI) and have adequate cooling to endure lots of heavy loads. I add this because it's not clear if you are looking to buy 10 new workstations and use them both as a cluster and as a compute resource, or if you are looking to get a dedicated computing cluster.

As for scheduler requirements, that is more driven by what you are looking to do. If you are running MATLAB only, then MDCS+Job manager will suit your needs fine. If you are looking to run other applications on the cluster or use sophisticated queuing techniques, then you might want to look into the third party ones, like HPC Server.

And I agree with Titus that more in-depth questions likely need to go to sales. There are a lot of requirements that should be identified before any money is spent to make sure the cluster does what you expect it to do.

[Danny]

thank you very much Jason. That computer cluster is dedicated for MATLAB only. The biggest problem is that in Hong Kong, there is only a reseller of matlab software. The reseller could hardly provide much technical support on this issue.

Regarding to the configuration of such cluster, computing node (i.e worker) & job manager to enable matlab programs running on them. Is it straightforward? I read the MDCS's installation guide, it seems straightforward. I am not sure in real implementation, this may not be the case. Any experience you could share on this front.

many thx

[Titus Edelhofer]

Hi Danny,

the installation is indeed straight forward (as long as you use the jobmanager as scheduler), you just follow the steps of the installation guide one by one ...

Titus

[Danny]

thank you very much for your sharing