Hi Atul, We have included Intel Xeon Phi Support for Univa Grid Engine in the year 2012 in Univa Grid Engine 8.1.3. For that it required to add some new functionality which was missing in Sun Grid Engine.
So basically what we did was: - Create a new resource type which allows you to do a mapping between a job and the selected MIC. This is needed in order to manage the mapping easily. - Adding support so that a MIC binary can be submitted directly to the cluster (which is then executed on a free MIC). - Added an customizable load sensor which allows you to monitor and account the state of the MICs (power consumption, different temperatures, etc.) - Added a command line tool with similar functionality - Added support for hybrid application which run partially on the normal Intel processor and offloads some function calls to the Intel Xeon Phi. This support includes that a job is automatically bound to the socket on which the selected Intel Xeon Phi card is attached to with the PCIexpress interface. Running an execd directly on the Intel Xeon Phi is not what we wanted todo since in some cases people want to restart the the co-processors after a job run / like on GPUs in order to reap the memory for privacy (just one reason). When it comes to KNL then things might change. With kind regards and sorry for posting this on our open source mailing list. But I think it could be from general interest. Daniel Am 24.07.2014 um 11:47 schrieb Atul Yadav <atulyadavt...@gmail.com>: > Dear Team, > > Thank you for providing input. > Then share the steps to represent the MICs and the prolog manipulates lock > files to assign > specific MICs to the job when it starts. > > In our setup we are having 4 servers with 2 mic installed on each server. > > Thank You > Atul Yadav > > ---------- Forwarded message ---------- > From: William Hay <w....@ucl.ac.uk> > Date: Thu, Jul 24, 2014 at 1:23 PM > Subject: Re: [gridengine users] Intel MIC integration with scheduler > To: users@gridengine.org > > > On Wed, 23 Jul 2014 17:18:09 +0000 > Atul Yadav <atulyadavt...@gmail.com> wrote: > > > Dear Team, > > > > We are planning to use gridengine scheduler for managing Intel MIC > > resources. > > > > We need your assistance building GE2011.11p1 with Intel MIC. > > > > Please share the steps for completing the activity. > I'm not sure there is such a thing as "the steps". We have several > machines with two MICs each and are still using an older version of > Grid Engine. We treat them as coprocessors. We have a complex defined > to represent the MICs and the prolog manipulates lock files to assign > specific MICs to the job when it starts. We then drop a file listing the > MIC to use in the job's TMPDIR in a similar fashion to the way parallel > environments create a machines file. At the moment we don't enforce > the allocation but rely on the honor system. > > An alternative to treating them as coprocessors would be to treat them > as execution hosts in their own right. Compile Grid Engine for the MIC > and run an execd on them(warning I've never tried this). If you wanted > to use them as a co-processor in this scenario you could then define PEs > with a single slot on the host and the remaining slots on the MICs. Use > -masterq to ensure the job starts on the host with SLAVE slots on the > MICs. > > William > > > > > _______________________________________________ > users mailing list > users@gridengine.org > https://gridengine.org/mailman/listinfo/users > > > <signature.asc>_______________________________________________ > users mailing list > users@gridengine.org > https://gridengine.org/mailman/listinfo/users
_______________________________________________ users mailing list users@gridengine.org https://gridengine.org/mailman/listinfo/users
