From Engineering Grid Wiki
Resource requests are flags added to the job submission that will cause SGE to look for specific CPU types, local free disk, free RAM, or other things.
IMPORTANT: Using these flags may cause your jobs to not run immediately, and sit in the queue until such time as the requested resources are free.
It's used as such:
These two flags control resource requirements. -soft means the requested resources would be nice to have, but not necessary. -hard indicates the resources are required. You can intersperse these flags amongst resource requirement flags - the first preceding option will define whether the following resource requirement is required or not.
This flag requests resources. Some common options are:
-l mem_free=XXG -l mem_total=XXG
The two flags above request certain amounts of memory. It's usually best to request mem_free, as your job will find the optimum free memory rather than a possible total.
This flag, unique to the Cloud, specifies the amount of local disk space on a node your job might use. It is your responsibility to move any data to and from this space, either as part of the job script or by hand. Data in this space is deleted after 30 days. The path your jobs should use to access this space is /cluster-tmp - this path is local to each node and is not shared, and not backed up. We recommend that any job requiring heavy disk access make use of the local temp space.
-pe smp4 4
This flag requests a 4 cpu smp machine, smp8 8 and smp2 2 also work. Since all machines are smp2, smp2 jobs have a better chance of getting a pair of CPUs to run, than smp8 or smp4 jobs, smp4 jobs will run before smp8 job. MATLAB will automatically use extra CPUs in large matrix operations, so ask for a machine that has them available.
-l cpu_type=intel -l cpu_type=amd
This flag requests a specific CPU type - AMD or Intel. Do note there are only 16 Intel CPUs available at this time.
This flag requests a certain "tier" of CPU - this allows you to request a specific class/architecture of CPU. Below is a table showing the current tiers and what sort of CPUs you'd expect in that tier per manufacturer.
|Tier 1||Tier 2||Tier 3||Tier 4||Tier 5|
|Intel||Xeon 53++||Xeon 56++||Xeon E5-2650|
|AMD||Opteron 8xx||Opteron 22xx/82xx||Opteron 23xx||Opteron 61xx||Opteron 62xx|
NOTE: These tiers are subject to change as hardware rotates through - there may be more tiers added, or CPU types may shift down a tier as the lower end equipment is retired.
Matlab - When less is more If your going to run a multi threaded matlab job you may want to test the number of threads to use, lets look at the following example:
t = int32(16); a = rand( 10000, 10000); while t > 1 maxNumCompThreads(t); fprintf( 'Threads: %d\n', t) tic;b=a*a;toc t = t / 2; end
Your mileage may vary