GPU1 and GPU2 available

[Predrag Punosevac]

Dear Autonians,

The members of the team who are occasionally checking our monitoring
tools or documentation probably have noticed that the Auton Lab has two
fully functional GPU computing nodes. This is the official announcement.
Both machine have identical specifications.

2xXeon E5-4627 2.4 GHz (6 Core+Hyperthreading)=24CPU in total
128 GB of RAM
1xTesla K80 with 4992 cuda processors

running Springdale 7.1 (Princeton clone of RedHat) and whole slue of
scientific software.

gpu1.int.autonlab.org was purchased from Dr. Barnabas Poczos grants
while gpu2.int.autonlab.org was purchased jointly by Dr. Artur Dubrawski
and Dr. Martial Hebert.

Dr. Poczos have kindly granted access to all members of the lab (members
of Neill group included) to his GPU computing node for at least 4 weeks.
I have created a scratch directories for all of you on 

/home/scratch/your_username 

Total size of scratch directory is 2 TB. I have not had a chance to
check with Dr. Dubrawski today if there are any restrictions on the use
of gpu2.int.autonlab.org so for now I have mounted /zfsauton/home shares
to gpu2.int.autonlab.org which means that you have the access to gpu2 if
you are the member of Dr. Dubrawski team.  His server also has 2TB of
scratch space


gpu[1-2}.int.autonlab.org accessible like any other computing node in
the lab.


I have configured CUDA drivers and I have done some light testing as
this print outs show

ptxas info    : 'device-function-maxrregcount' is a BETA feature
(target: sm_52)
mkdir -p ../../bin/x86_64/linux/release
cp simpleDevLibCUBLAS ../../bin/x86_64/linux/release
make[1]: Leaving directory
`/usr/local/cuda-7.5/samples/7_CUDALibraries/simpleDevLibCUBLAS'
Finished building CUDA samples


>> gpuDevice


ans =
       
  CUDADevice with properties:
       
                      Name: 'Tesla K80'
                     Index: 1
         ComputeCapability: '3.7'
            SupportsDouble: 1
             DriverVersion: 7.5000
            ToolkitVersion: 7
        MaxThreadsPerBlock: 1024
          MaxShmemPerBlock: 49152
        MaxThreadBlockSize: [1024 1024 64]
               MaxGridSize: [2.1475e+09 65535 65535]
                 SIMDWidth: 32
               TotalMemory: 1.2079e+10
           AvailableMemory: 1.1939e+10
       MultiprocessorCount: 13
              ClockRateKHz: 823500
               ComputeMode: 'Default'
      GPUOverlapsTransfers: 1
    KernelExecutionTimeout: 0
          CanMapHostMemory: 1
           DeviceSupported: 1
            DeviceSelected: 1

I have also played little bit with PyCUDA. R is also capable of
utilizing GPU modules but I have not played with it. 



Finally both servers have additional 10TB storage space. Permission for
use of that space will have to be granted by Dr. Poszos and Dr.
Dubrawski personally. Finally these machines are specifically purchased
for GPU computing. Any jobs which don't involve GPUs will be
indiscriminately terminated by me.

Best,
Predrag