 |
 |
I can speak only for atlas they are aiming also at 8cores but if they decide for the whole node they are going to use it or the unused CPU needs to be accounted for as if it was used. I know nobody asked yet. cheers alessandra On 19/09/2013 11:42, Andrew Lahiff wrote: > > Hi, > > Didn't the TEG decide on "multicore" rather than "whole-node"? > Whenever I've asked others in CMS about what they want the answer has > always been "multicore". > > In our case if a job requests 8 cores and ends up on a 64 core worker > node, the job will only use 8 slots and the remaining slots will be > used by other jobs. We could easily allow for whole node jobs as well, > but so far no VOs have asked for this. > > Regards, > > Andrew. > > *From:*Testbed Support for GridPP member institutes > [mailto:[log in to unmask]] *On Behalf Of *Sam Skipsey > *Sent:* 19 September 2013 11:22 > *To:* [log in to unmask] > *Subject:* Re: Technical Meetings > > On 19 September 2013 10:48, Christopher J. Walker > <[log in to unmask] <mailto:[log in to unmask]>> wrote: > > On 19/09/13 10:12, Andrew Lahiff wrote: > > Hi, > > For the record, for the new batch system at RAL we have multicore queues > on our CREAM CEs (currently configured to use 8 cores). However, on the > ARC CEs jobs can request exactly how many cores (and how much memory) > they need rather than having to use a specific queue. > > I believe it is perfectly possible to do this with CREAM too (though > I'm not sure we have it set up on all CEs. > > Sure, it's basically how our MPI support works at Glasgow, IIRC, and > we're entirely behind CREAM CEs. > > IIRC, when I asked the experiments why this wasn't sufficient and > they wanted to auto discover how big the slot they had got was, it > was because on a 12 core node, if you have jobs requesting 8 > slots, they may actually end up on a 12 slot machine - and if they > know this they can make use of the extra slots they discover they > have. > > Well, this is the difference between "whole node queues" and > "multicore queues" (and between shared memory multicore queues and > message passing multicore queues). Our MPI support provides precisely > that - so a job can request any number of cores and it'll get them, > but almost never all on the same node, as they don't need to be for MPI. > > (This seems sufficient for biomed, and indeed any other entity that > writes MPI-based code.) > > We don't support OpenMP style shared-memory parallelism where you > require N slots all on the same node. (This makes the scheduling > problem harder as has been discussed before.) > > ATLAS/CMS seem to want whole node queues, in which case, if you assume > a whole node queue is sensible a priori, it is reasonable for them to > want to know how big the node they'll get is in advance. > > (This would be particularly pessimal for an 8 slot job arriving on a > 64 slot node, for example.) > > We did have a whole node queue at Glasgow for testing (Andy Washbrook > used this), but the scheduling was exceedingly pessimal (as running 10 > jobs would offline 10 nodes, without checking how big the nodes were > first... so if it hit a 64 core node...) so we turned it off. > > Sam > > This is how both > ATLAS and CMS are running multicore jobs at RAL now. Condor is > then > responsible for scheduling the mix of single and multicore jobs. > > Chris > > > Regards, > Andrew. > > -----Original Message----- > From: Testbed Support for GridPP member institutes > [mailto:[log in to unmask] > <mailto:[log in to unmask]>] On Behalf Of Alessandra Forti > Sent: 19 September 2013 09:57 > To: [log in to unmask] <mailto:[log in to unmask]> > Subject: Re: Technical Meetings > > Well that's for testing. If it goes in production it will be > more than one node and most of the things that are keeping > this back are operational i.e. how not to waste resources and > how to do the accounting if a job requests a certain number of > CPU or the whole node. > > cheers > alessandra > > On 18/09/2013 18:18, Christopher J. Walker wrote: > > On 18/09/13 17:22, David Colling wrote: > > Hi Alessandra, > > Yes they are indeeedee. I only know of bits and pieces in the LHC > world but do know, for example, that our T2K colleagues make > extensive use of them. The Imperial T2K people code and debug > locally > and then run on the RAL resources. This is proving so > successful that > we are considering adding a bigger node - perhaps to the GridPP > cloud so that others could use it via OpenStack. These are at the > *ideas* stage at the moment, but if we did would there be any > takers > or would we have just thrown away a chunk of money (or rather > given > it to T2K as I am sure that they would use them)? > > I guess that the question is what should GridPP be doing about > this? > I don't see it as our place to fund development in the individual > experiments but should we be acting as a conduit for best > practise? > Organising Goofit tutorials? Interacting with EGI as Stephen > suggests? > What else? Is there a focus that we can develop with very > little money? > > I think that these are questions for next Tuesday rather than > Friday > but I will add a specific discussion to the discussion agenda > for this. > > QMUL now has a single node MPI queue - Dan has more details. > What more > does one need? > > Chris > > Best, > david > > On 18/09/13 14:41, Alessandra Forti wrote: > > Hi, > > multicore should become a reality at the end of LS1. We should > definitely have it as an activity. > > cheers > alessandra > > On 18/09/2013 14:12, David Colling wrote: > > Hi, > > I am just drawing up an agenda and is it worth having an item on > FTS3 (from Andrew L.)? > > Also the many and multicore activity. Is somebody able to describe > what has been happening in these two areas? Is this something that > we want to have as an activity in GridPP5? > > Best, > david > > > > -- > Facts aren't facts if they come from the wrong people. (Paul > Krugman) > > > -- > Scanned by iCritical. > > -- Facts aren't facts if they come from the wrong people. (Paul Krugman)