You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 18 Next »

The Shared (General) Farm consists of several physical clusters that are available to all SLAC users. The cluster hardware was purchased incrementally over several years by various stakeholder groups. Each physical cluster is based on a specific hardware model but all hosts in the farm run 64bit RHEL/CentOS6. The LSF "general" queues feed user's jobs to the shared farm. The stakeholders have associated LSF user groups with a fairshare (scheduling priority) that reflects their cluster investment. This ensures that stakeholders will always get some runtime on the cluster when utilization is high. Users that are not members of stakeholder groups still have the ability to run jobs "for free". There is a superset fairshare group "AllUsers" that includes all SLAC users. A non-stakeholder must compete for priority with all other users running jobs on the shared farm. This may be acceptable for some, but production environments may demand priority scheduling. The free "AllUsers" fairshare is actually subsided by the paying stakeholders. A stakeholder's fairshare value is derived from the compute power they have purchased. A HS06 CPU benchmark is calculated for each cluster server model.

 

Stakeholder Investments

 

Cluster# of hosts# cores/host# of cores# Batch slotsHS06/slotHS06Owner/GroupPurchased dateNotes
fell - - - - - - -2007 IGNORE - obsolete hardware in run-to-fail mode
hequ39831231215.014683.12ATLAS9/23/2009 
hequ76860860815.019126.08Fermi9/23/2009 
hequ 77861661615.019246.16Babar9/23/2009 
dole381245645613.776279.12ATLAS10/19/2010 
kiso6824163213609.9713559.2ATLAS9/23/2011HT enabled/ 20 slots per host
bullet77.5161240124016.0519902PPA 12/5/2012for MPI use - do not map to fairshare
bullet76.25161220122016.0519581Fermi12/5/2012 
bullet17.751628428416.054558.2Geant12/5/2012 
bullet47.251675675616.0512133.8ATLAS2/13/2014 
bullet49.251678878816.0512647.4PPA2/13/2014 
bullet19.51631231216.055007.6Theory2/13/2014 
bullet32.51652052016.058346Beamphysics10/2/2014for MPI use - do not map to fairshare
deft222452852814.977904ATLAS1/1/2016HT disabled, two purchases at 11/2015 and 1/2016
deft72416816814.972514.96Fermi6/1/2016 

bullet MPI users

The bulletmpi queues are not associated with these fairshare policies. MPI parallel jobs reserve cores across multiple hosts. This does not work well with fairshares because the fairshare formula takes reserved cores into consideration. Large parallel jobs would never start because a job's dynamic scheduling priority would drop before it reserves all the cores it needs to get started. The bulletmpi queues have a higher queue-level priority than all of the fairshare general queues. For this reason we do not assign general queue fairshares for MPI investments.

AllUsers 'Tax'

The stakeholders subside the AllUsers ("free") fairshare by paying a 15% tax on their shares: 

GroupHS06HS06 (after tax)
ATLAS4455937875
Fermi3122226539
Babar92467859
Geant45583874
Theory50084257
PPA others1264710750

The 15% tax results in 16086 shares for AllUsers

LSF fairshare user groups

The stakeholders are responsible for distributing their shares among their respective LSF groups. The table below should match the production LSF config:

USER/GROUPHS06_SHARES%HS06_SHARESHS06_OWNER
    
atlasgrp3787535.32%ATLAS
babarAll78597.33%BaBar
glastdata8540.80%Fermi
glastusers2532023.61%Fermi
glastgrp3660.34%Fermi
geantgrp38743.61%Geant
luxlz35003.26%PPA others
cdmsdata20001.86%PPA others
lcdprodgrp11001.03%PPA others
exoprodgrp15001.40%PPA others
hpsprodgrp10000.93%PPA others
rpgrp5000.47%PPA others
lcd6000.56%PPA others
exousergrp5500.51%PPA others
rdgrp00.00%PPA others
theorygrp42573.97%Theory
All Users1608615.00%Everyone (15% tax)

Decommissioning clusters

Fairshares should be subtracted from stakeholder groups and the associated AllUsers tax when hardware is declared obsolete and put into run-to-fail mode. The fairshare distribution should always reflect stakeholder investment in current production cluster resources.

Chargeback model for sustainable hardware lifecycle

This effort paves the way for a possible chargeback model where Computing Division could lease cluster hardware and charge stakeholders a rate for fairshares.

 

 

 

  • No labels