UP Site Selector for Grid03

1
UP Site Selector for Grid03

• Catalin Dumitrescu, Mike Wilde, Ian Foster
• The University of Chicago Argonne National
  Laboratory

2
Introduction

• Problem Usage Policy (UP) based scheduling in
  Grid03
• Contention problems are an issue when
• No resources are free
• Restrictions/Quotas are enforced at sites
• !-------------------------------------------------
  -------------------------------------!
• ! Site Name ! ! iVDGL ! USAtlas !
  Ligo ! SDSS ! bTEV ! Gridex ! USCMS !
• !-------------------------------------------------
  -------------------------------------!
• cluster28.knu.ac.kr 71 16.17 16.17
  16.17 16.17 16.17 1.06 1.91
• garlic.hep.wisc.edu 101 3.01 3.01
  3.01 3.01 3.01 - 1.43
• nest.phys.uwm.edu 305 0.00 0.00
  7.28 0.00 0.00 - 0.00
• t2cms0.sdsc.edu 76 0.62 24.74
  - 24.74 24.74 0.00 0.40
• uscmstb0.ucsd.edu 3 11.68 11.68
  - - - 11.68 11.68
• xena.hamptonu.edu 1 25.00 25.00
  - - - - 25.00
• !-------------------------------------------------
  -------------------------------------!
• Our Solution an UP site selector for steering
  workloads in Grid03

3
Enhancement

• Currently
• Site Name Class Location VO Administrator
  CPUs
• Our proposal
• Site Name Class Location VO CPUs
  Allocated to Grid03 Current UP per VO Usages
  per VO

4
Architecture Overview
Site Selector
select site whererecommendation-honored and
condition
Start-time Predictor
Add start-time prediction to Choice-Table
Site Recommender (can be multi-level Grid, VO,
and Group)
Job description
UP DB
SiteRec (VO, ResRequest) gt Choice-Table
5
UP Example
Resource allocation cases UP quota
Under-allocation due to external causes
Over-allocation without contention
Over-allocation with contention
Under-allocation due to internal causes
6
UP Metrics Flow

• Planner, Work-Runner, or Application
  infrastructure invokes V-PEP to get a recommended
  execution site for each job to run. This
  recommendation is policy-cognizant. (site and VO
  policies)
• V-PEP
• provides answers based on site policies, VO
  policies, Job requirements, sampled monitoring
  data
• gets queue conditions, local site policy from
  monitoring system gets job resource requirements
  from job declaration and predictor data
• There can be one V-PEP per VO or one V-PEP per
  Grid, and V-PEPs can be multi-layer (Grid, VO,
  and Group)
• The place to do resource reservation, especially
  storage, needs to be worked out many
  architectural alternatives

7
MAUI-like UP (semantic)

• Goal fair-share resource allocation among users
• Provides support for fair-share rule
  specification
• Example bettysue 15.5, billybob 10.0, jimbob
  5.0-
• MAUIs formula
• JobPriority (Other Priority Components)
  FairShareFactor
• FairShareFactor FairShareWeight
  ((UserFairShareWeight UserFairShareDelta)
  ((GroupFairShareWeight GroupFairShareDelta)
  ((AccountFairShareWeight AccountFairShareDelta))
  
• Fair share is computed on intervals with DECAY
  factors
• http//www.hpc2n.umu.se/doc/maui/fairshare.html

8
UP Selector Service

• Specialized Grid Service (OGSI)
• Collects Abstract Usage Policies from sites
• Queries the Grid03 monitoring systems for various
  metrics
• Answers queries like
• From the list L of sites, which is the subset S
  of sites where VO Vs workload is allowed to
  run?
• Which is the best site X to send VO V's
  workload?

9
Input Metrics / Decisions

• Input Metrics
• CPU capacities and usages of CPUs, available
  CPUs, used CPUs (totally and per VO), allocated
  to Grid03 and per VO
• Free and used disk spaces (APP, DATA directories,
  quotas)
• Network availability (latency, capacity)
• Decision Examples
• select best site for CMS
• list all sites available for CMS and their
  characteristics

10
UP Selector Text Client

• Used for interfacing with different other user
  schedulers
• Accepts simple commands such as
• List sites with characteristics
• Select best
• Commit usage

11
Usage Example Euryale

• Nice shell and perl wrapper of the java client
  
• ./run-client.sh (before anything else) - window1
• Provides in addition local access to enhanced
  selection
• Euryale style request
• ./mock-up /tmp/kss-gj4CUt.lof (window2)
• SOLUTIONRice_Grid3

12
UP Selector Graphic Client
13
Conclusions

• The problem we are taking on Usage Policy (UP)
  based scheduling in Grid03
• Our solution an UP site selector for steering
  workloads in Grid03

14
References

• S-POP.v3.0 http//people.cs.uchicago.edu/cldumi
  tr/Grid03_policy/S-POP/v3.0/
• V-PEP.v3.0 http//people.cs.uchicago.edu/cldumi
  tr/Grid03_policy/V-PEP/v3.0/
• S-Console http//people.cs.uchicago.edu/cldumit
  r/Grid03_policy/S-Console/
• S-POP.data http//griodine.uchicago.edu/cldumit
  r/data/
• Grid03_policy http//people.cs.uchicago.edu/cld
  umitr/Grid03_policy/