Client struggles to get 256 cores of MDCS

Having just installed the MDCS with 256 cores, I'm finding that my clients just can't pick up the full host.

I have a single MJS queue with all the cores in it.

The Validation passes all tests except the final Parpool test.

I think I have sufficient ports open - TCP, 27000-29150

(base port = 27350, 1000 bc parallel, 5, nJ =1, nW = 256, 3nw=768; bp+1000+5+nj+3nW = 29124, I rounded up for good measure)

I can get the client to work with 128 cores, but it just wont get 256.

Answer 1

[Lachlan Simpson]

This was solved by addressing the File Descriptor issue. Unfortunately, again, MatLab's documentation is not only poor, but uninformative.

When I tried ulimit in Ubuntu 14.04 I got "unlimited" as both root and user.

Turns out "ulimit -a" gave a richer answer.

I used a combination of this post wrt hard coding limits and this post to discover that ulimit is quite rich and the man page is quite poor . (hint: ulimit -a tells everything)