My matlab mdcs worker slower

hi

I tried my code on local config it work fine but when I'm using the cluster it work slower every time i increase the number of worker the time rise

the cluster is rocks with one head node and 3 compute node

in installation i install one copy for head node and another copy for compute node shared with nfs using mjs and i need to use large data set of images can i shared before i start the matlab pool

thanks

Answer 1

[Jason Ross]

Keep in mind the following:

1. When you are using the "local" cluster, everything is local to the machine. So you are using local disk for the workers and (possibly) the data. That's going to be pretty quick.
2. When you move to starting and running MATLAB from the NFS server, the application starts and runs over NFS. NFS can be pretty good at this, using caching and other tricks to make it better -- but it's still not the same as running from local disks in most cases, startup will likely be slower, as will accessing "new" things.
3. If you are using large data sets as well as NFS to run the workers, this can also slow processing time. Especially if you are trying to move lots of small files around, this will also affect performance adversely.
4. You also need to keep an eye on your NFS server's load as well as the bandwidth utilization of your network ports. It's possible that running MATLAB and trying to use the network to move the data around is really hurting you.

I'd recommend trying to use local installs of MATLAB on all the cluster machines and only use the network for getting the images down, since I'm betting that it's far more convienient for your design to be able to access the network data, but it's not adding much value to run the application over the network.

Also, you might want to experiment with the number of workers. It's possible that the workers are overwhelming either the processor or network connection of your network server, and generally when NFS is overwhelmed, the response curve is very much the "hockey stick" graph ... response times can go from sub-millisecond to double digit millisecond very quickly, and that just kills performance. So it's not unknown that you would do more with a smaller number of workers.