This is based on Son of Grid Engine 8.1.9. Added suppotrs for GPU scheduling, works for searial, paralle, MPI, array jobs.(Check this repository too: https://github.com/prod-feng/sge-gpu/tree/master)

Supports seting up hybriding GPU and CPU jobs on the same nodes for resource sharing between users at the same time.

Compile tested using the followinf commands:

#go to the source folder

cd $SGE_CODE/source

./aimk -no-java -no-jni -no-secure -spool-classic -only-depend -no-qmon -debug

./scripts/zerodepend

./aimk -no-java -no-jni -no-secure -spool-classic  -no-qmon -debug depend

./aimk -no-java -no-jni -no-secure -spool-classic  -no-qmon -debug

#to install:

export SGE_ROOT=/where/you/want

scripts/distinst -all -local -noexit
(install both sgemaster and execd, good foe testing)

#debug:

source ./dist/util/dl.sh
 
dl 10

(then restart sgemaster, and/or execd services)

(also see here: https://github.com/prod-feng/sge-gpu/tree/master)

Simultaneously run CPU jobs and GPUs jobs on the same node, one would better to reserve 1 CPU core per 1 GPU. The way I do is:(suppose nodeA has 40CPU core, 2 GPUs) (there should be other ways)

1. Define a "rcpus" comsumable:

qconf -sc|grep rcpu
rcpus               rcpus      INT         <=    YES         YES        10       0

Here default debit is "10".

2. Set the value of the "rcpus" as:

40 cores - 2 cores(for GPU) =38 X 10 =380

rcpus=380+2 =382

3. Update the node

qconf -se nodeA
...
complex_values        slots=40,rcpus=382,ngpus=2
...

4. Name CPU queues as "cpu*", and GPU queues as "gpu*'
5. Add JSV to SGE conf

qconf -sconf

...
jsv_url                      /somewhere/mygpujsv.pl
...

In this "mygpujsv.pl" file, add the codes:

        if ( $key=~/gpu/) {
          jsv_sub_add_param('l_hard', 'rcpus',1);
        }
.......
(or add the following to make sure)

        if ( $key=~/cpu/) {
          jsv_sub_add_param('l_hard', 'rcpus',10);
        }

In this way, CPU jobs will debit 10 "rcpus", which can only use 38 cores; while, GPU jobs will debit 1 "rcpus". So reservs 2 CPU cores for GPU jobs.

====================================

-pe openmp 10    # requests 10 cpu cores

-l ngpus=0.2     # so here 10X0.2=2 GPUs for this job on the same node.

listener_threads       1
worker_threads         1

NO Guarantee!

======================================

Add patched_files_ge2011.11.p1.0.1.tar.gz .

Fix a bug to support GPU array jobs properly. Only for GE2011.p1 now.

NO Guarantee!

======================================

======================================

This is a package which is designed to enable GPU scheduling capability to GE2011.11p1(not applicable to other versions yet). With this patch, SGE can support GPU without any external Wraper tools.

Patch to Son of Grid Engine 8.18 is available too in the following link.

$ qconf -sc

#name               shortcut   type        relop requestable consumable default  urgency 
 ...
ngpus               gpu        INT         <=    YES         YES        0        5000

qconf -se node1
...
complex_values        slots=12,ngpus=2,...

qsub -l ngpus=1 ...

This also works for parallel jobs.

qsub -pe openmpi 4 -l ngpus=1 ...

It supports multiple GPU scheduling on multiple nodes for parallel jobs(MPI, etc.) as well.

For example, if node1 and node2 each has 4 GPUs installed.

On node1, JobA uses GPU0, JobB uses GPU2;

On node2, jobC uses GPU 1 and GPU 2.

And then JobZ requestes 4 GPUs, the patched SGE can dispatch GPU1 and GPU3 on node1, GPU0 and GPU3 on node2 to JobZ, and set the environment for the job on node1 as:

CUDA_VISIBLE_DEVICES=1,3  (0,2 are alreadt used by jobA and jobB)

and environment for the same job on node2 as:

CUDA_VISIBLE_DEVICES=0,3  (1,2 are already used by jobC)

For non-GPU jobs, CUDA_VISIBLE_DEVICES is set to be empty.

With this patch, you do not need wrapper or load sensor anymore to schedule multiple GPU jobs.

Check the link too: http://sourceforge.net/projects/ge-gpu/?source=directory