\\
**[[cluster:0|Back]]**

==== HPC Users Meeting ====

  * Brief history 
    * 2006 swallowtail (Dell PE1955, Infiniband, imw, emw)
    * 2010 greentail (HP gen6 blade servers, hp12)
    * 2013 sharptail (Microway storage, K20s, Infiniband, mw256/mwgpu) 
    * 2014 mw256fd (Dell 2006 replacement with Supermicro nodes)
    * 2015 tinymem (Supermicro bare metal, expansion for serial jobs)
    * 2017 mw128 (first new faculty startup funds)
    * 2018 6/25 Today's meeting

  * Since 2006
    * Grown from 256 to roughly 1,200 physical CPU cores
    * Processed 3,165,752 jobs (by 18jun2018)
    * Compute capacity over 60 teraflops (DPFP; 38 cpu side, 25 gpu side)
    * Total memory footprint is near 7.5 TB
    * About 500 accounts have been created (incl 22 collaborator and 100 class accounts)
  
  * Funding / charge scheme: is it working for you?
      * Last 2 years, $15K target realized each year.

  * Status of our cluster development fund
    * $140K come July 1st, 2018
    * Time for some new hardware?  Retirement of hp12 nodes?
 
  * 2017 Benchmarks of some new hardware
    * Donation led to purchase of a commercial grade GPU server containing four GTX1080ti GPUs
    * Amber 16. Nucleosome bench runs 4.5x faster than on a K20
    * Gromacs 5.1.4. Colin's multidir bench runs about 2x faster than on a K20
    * Lammps 11Aug17. Colloid example runs about 11x faster than on a K20
    * FSL 5.0.10. BFT bedpostx tests run 16x faster on CPU, a whopping 118x faster on GPU vs CPU.
    * Price of 128gb node in 2017 was $8,250...price of 256gb node in 2018 is $10,500

  * 2016 IBM bought Platform Inc (developers of LSF, Openlava is LSF4.2 open source branch)
    * IBM promptly accused Openlava of copyright infringement in v3/v4 (US DMCA law, no proof needed).
    * Fall back option to v2.2 (definitely free of infringement, minor disruption)
    * Move forward option, adopt SLURM (LBL developers, major disruption)

  * If we adopt SLURM should we transition to OpenHPC Warewulf/SLURM recipe?
    * http://openhpc.community/
    * new login node and couple compute nodes to start?

  * New HPC Advisory Group Member

  * Tidbits
    * Bought deep U42 rack with AC cooling onboard and two PDUs
    * Pushed Angstrom rack (bss24) out of our area, ready to recycle that (Done. 06/20/2018)
    * Currently we have two U42 racks empty with power 
    * Cooling needs to be provided with any new major purchases (provost, ITS, HPC?)
    * 60 TB raw storage purchased for sharptail (/home2 for users with specific needs)
    * Everything is out of warranty but
      * cottontail (03/2019), 
      * ringtail & n78 (10/2020)
      * mw128_nodes & shartptaildr (06/2020)
    * All Infiniband ports are in use

===== Notes =====

  * First make a page comparing CPU vs GPU usage which may influence future purchase [[cluster:167|CPU vs GPU]]
  * $100k quote, 3to5 vendors, data points mid-2018
  * One node (or all) should have configured on it: amber, gromacs, laamps, namd, latest version
  * Nvidia latest version, optimal configs cpu:gpu ratios
    * Amber 1:1 (may be 1:2 in future releases) - amber certified GPU!
    * Gromacs 10:1 (could ramp up to claiming all resources per node)
    * Namd 13:1 (could ramp up to claiming all resources per node)
    * Lammps 2-4:1
  * 128g with enough CPU slots to take over ''hp12'': double ten core minimum
  * Anticipated target (also to manage heat exchange)
    * 2x10 Xeon CPU (~100gb left) with 2xgtx1080ti GPU (25gb memory required)
    * as many as fit budget but but no more than 15 rack wise

\\
**[[cluster:0|Back]]**