We have several compute servers that are dedicated to the machine learning group.

  • henwen -- dual Xeon circa 2012, 64GB RAM, 4TB local disk
  • wilbur -- dual Xeon circa 2012, 64GB RAM, 4TB local disk
  • enceladus -- dual 2.66GHz Xeon, 32GB RAM
  • titan -- single 3.0 GHz Xeon, 3GB RAM
  • neptune -- single 3.0 GHz Xeon, 3GB RAM
  • pluto -- single 3.2 GHz Xeon, 3GB RAM
  • mercury -- dual 3.2 GHz Xeon, 4GB RAM
  • venus -- dual 3.2 GHz Xeon, 4GB RAM

Also there are two servers that are reserved for GPU users. Access to GPU servers should be discussed with Amos Storkey. Use of the servers involves informal collaborative management. Krzysztof Geras is the currentGPU machine manager and you should discuss your scheduling requirements with him.

  • melmac -- single Xeon 3.7 GHz, 64GB RAM, 2 x GTX 780, 2014
  • roswell -- single i7 2.8 GHz, 6GB RAM, 2 x GTX 480, c 2011

Finally there is stonesoup, an ANC GPU machine with 4 Tesla K40 boards circa Dec 2014. The tesla cards are dedicated to particular projects, but the machine as a whole can be accessed by other researchers. We suggest this machine ls reserved for use with production runs.

When using these always check if it is not busy with /disk/scratch/imurray2/gpu_lock on roswell and /disk/scratch/gpu_lock on melmac.

I have placed these in order of assumed desirability, best servers first. If you have a different opinion feel free to reorder the list.

(NB I essentially guessed this info based on /proc/cpuinfo, etc. These servers were acquired over a number of years, and the Xeon platform has changed substantially in that time, so some "Xeon"s will be significantly faster than others, regardless of stated CPU speed. It appears that all of the servers we have are quad core. Feel free to add other notes about the individual servers.)

GPUs

If you're going to use GPUs (on roswell), please join ancmlgpu-users@infREMOVE_THIS.ed.ac.uk and use the discretionary locking system described at http://homepages.inf.ed.ac.uk/imurray2/code/gpu_monitoring/

Please refrain from CPU-only jobs on roswell to save the machine for the GPU users.

Storage

We have storage available in several places. Please let one of the faculty know if they fill up, so that we can talk about getting more.

  • Backed-up AFS disk space is available in /afs/inf.ed.ac.uk/group/ANC/
  • There is also /afs/inf.ed.ac.uk/group/ML which is NOT backed up, but it is mirrored to KB, so there is some disaster recovery.
  • Finally on each of the compute servers look in /disk for local, fast, but non-backed up scratch space.

Access Control

You may need to contact support in order to get permission to log in to these machines. You need to be a member of the anccs-user role, except for roswell, in which case you need the ancmlgpu role. The commands roles and getrolemembers may be useful. Probably, however, you should just tell support what machine you want to log into and let them figure it out.

-- CharlesSutton - 21 Feb 2012

Topic revision: r8 - 27 Feb 2016 - 17:25:59 - Main.jfowkes
 
This site is powered by the TWiki collaboration platformCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback
This Wiki uses Cookies