Parallel Session B2 - CPU and Resource Allocation

1
Parallel Session B2 - CPU and Resource Allocation

• Panelists
• Charles Young (BaBar)
• David Bigagli
• Seed Questions
• Batch queuing system in use?
• Turnaround guarantees?
• Pre-allocation of resources?

2
Batch System

• Vendor or home-brewed?
• Maintenance and support issues.
• If vendor, licensing and cost issues.
• Already significant fraction of H/W.
• Per node? Per unit computing power? Or?
• Management concerns.
• Can one really management 10K nodes?
• Split into separate management domains?

Slide from Charles Young (BaBar)
3
LSF - Talk given by David Bigagli (Platform
Computing)

• What is LSF
• Developer's view of LSF
• architecture
• Scalability
• Dealing with resources
• Load Information manager
• Batch
• Lively discussion

4
Discussion which batch systems are you using and
why?

• LSF (30 ?)
• LSF has worked for us and continues to work.
• PBS (30)
• PBS is free
• Collaborators want us to use PBS (because PBS is
  free)
• Ability to modify source
• Nobody is using ProPBS
• Condor (20)
• Condor costs nothing
• Cycle stealing allows us to get computing done

5
Discussion which batch systems continued...

• BQS
• Homegrown at IN2P3
• used in a small number of external sites
• Have had it for seven years
• Everyone likes it
• FBS
• Homegrown at FNAL
• used in some external sites
• LSF is expensive
• FBS is designed to be used on farms
• Lightweight and flexible

6
Other issues

• Mosix
• Only CERN has looked at it
• Appears to be difficult to take down individual
  machines in the cluster
• How to deal with abusers?
• Turn them over to the user community
• How do people schedule downtime ?
• Train people that jobs longer than 24 hours are
  at risk.
• CERN posts a future shutdown time for the job
  starter (internal)
• BQS has this feature inside.
• Condor has eventd for draining. Labs reboot and
  have maintenance windows.