Demandbased Coscheduling of Parallel Jobs on Multiprogrammed Multiprocessors

1
Demand-based Coscheduling of Parallel Jobs on
Multiprogrammed Multiprocessors

• Review
• Presented by Nanyan Jiang

2
Motivation

• How to coschedule jobs if the number of processes
  larger than the number of processors
• Traditional coscheduling
• Cost of global synchronization for traditional
  coscheduling
• Overhead of context switching (alternate
  selection)

3
Demand-based coschduling

• Dynamic coscheduling was proposed as a way to
  reduce the amount of global synchronization of
  coscheduling, and thus decrease the systems
  overhead and improve its scalability.

4
Dynamic Coscheduling
5
Flaws of always scheduling

• Completely unfair and favor jobs that send a lot
  of messages
• Even if message-sending rates are equal, this
  algorithm may take a very long time to switch out
  of a state in which most processors are running
  one job, although this dynamic behavior of the
  algorithm cannot be seen from the steady-state
  probability alone.

6
equalizing dynamic coscheduling

• If system runs no more than H scheduler cycles
• When h -H, ? always switch
• When h H, ? never context-switch due to
  arriving messages
• Fairness vs. coscheduling
• Experiment 10,000 cycles, h -1000, fair
  performance, little coscheduling
• Failure of coscheduling under simple equalization
• Disregarding more opportunities to coschedule
  processes
• The higher the value of h, the less coscheduling
  was achieved.
• We ensure the scheduler to make progress from job
  to job

7
Conclusion

• Demand-based coscheduling promises better
  performance by coscheduling only those processes
  that communicate with each other
• The key factor in achieving good coscheduling
  behavior under dynamic coscheduling is the mean
  communication rate
• Demand-based coschueding is
• Non-intrusive programmer is not required to
  write parallel programs in a particular style.
• Flexible can schedule large number processes on
  limited number of nodes
• Dynamic communication is detected automatically
• Dencentralized scheduling decision made locally

8
Discussion
9
Applications and scheduling schemes
10
New application domains

• Coscheduling of shared resources in new emerging
  applications heterogeneous environments
• Scheduling streaming applications on multiple
  sites in p2p networks
• Any other applications?

11
Dynamic coscheduling vs. Implicit Scheduling