1. A Characterization of Approaches to Parrallel Job Scheduling
Scheduling of parallel jobs in a heterogeneous grid environment
Each site has a homogeneous cluster of processors, but processors at different sites have different speeds
Much of the research on scheduling in heterogeneous systems has focussed on independent sequential jobs
Research on parallel job scheduling has concentrated primarily on the homogeneous context
The algorithms used for scheduling sequential tasks on heterogeneous systems are too computationally complex to extend to parallel jobs
We extend the techniques used for parallel job scheduling in a homogeneous context to the heterogeneous context
Problem Addressed
A Characterization of Approaches to Parrallel Job Scheduling
Simulation Environment
Heterogeneous sites, with a homogeneous cluster of processors at each site
5000 job subset of either the 430 node Cornell Theory Center (CTC) trace or the 128 node IBM SP2 system at the San Diego Supercomputer Center (SDSC)
NAS Parallel Benchmarks 2.0 used to model the heterogeneous runtimes
24.2
35.6
94.893
20.328*
LU Class B
(256 Nodes)
1.1*
1.8
2.2724
1.3147
MG Class B
17.2*
25.3
34.3
35.5
(8 Nodes)
17.7
16.3*
22.6
23.3
IS Class B
IBM
SP (P2SC 160 MHz)
Cray
T3E 900
IBM SP
(WN/66)
SGI
Origin 2000
* Denotes best runtime for the job
Gerald Sabin Rajkumar Kettimuthu Arun Rajan P Sadayappan Supported in part by Sandia National Laboratory
Metrics
Conservative vs. Arrgessive
Jobs are processed in arrival order by the meta-scheduler
Greedy assigns each job to the site with the lowest instantaneous load
Greedy-MR (Multiple Requests) submits each job to all sites
When the job starts at a site, the other instances are removed
We have shown this mechanism to be effective in a homogenous context (HPDC ’02)
However, only a slight improvement is seen in a heterogeneous context
We use the following metrics for evaluating the proposed schemes
Average Slowdown
Average Turnaround Time
Utilization
Effective Utilization
When network bandwidth is limited, jobs can be submitted to a smaller number of sites and the multi-reservation scheduler can still realize a substantial fraction of the benefits achievable from a scheduler that schedules each job at all sites, by making fewer reservations
Use a more accurate approach to select which sites the job is submitted, instead of using the instantaneous load, we query the site to determine the earliest completion time
When using completion time as the criteria, submitting to fewer sites can be almost as effective as submitting to all sites
Restricted Multi-Site Reservations
Backfilling
Backfilling
A later arriving job is allowed to leap frog previously queued jobs
Aggressive vs. Conservative
Processors
Processors
Time
Time
Processors
Time
Jobs are processed in arrival order by the meta-scheduler
In a heterogeneous context, the site where the job starts the earliest may not be the best site
In order to get the completion of a job at a particular site, conservative backfilling has to be employed at the local site
Conservative performs better than aggressive in all case, quite the opposite of a homogenous context
Improved backfilling caused by holes created due to the dynamic removal of replicated jobs at each site, and an increased number of jobs to attempt to backfill at each site
Conservative
Every job is given a reservation when it enters the system and a job is allowed to backfill only if it does not violate any of the previous reservations.
EASY
Only the job at the head of the queue is given a reservation and a job is allowed to backfill if it does not violate this reservation
Explicitly take into account efficacy to improve the effective utilization
Use efficacy as the priority order for the jobs in the reserved and idle queue
Starvation free
Effective utilization increases and turn around time decreases, in spite of the decreases in raw utilization
Efficacy Based Queues
Conclusions and Future Work
Improvement in turn around time and effective utilization for parallel job scheduling in a heterogeneous grid environment have been demonstrated through simulation
Next Steps:
Incorporate these changes into the Silver/Maui Scheduler
Deploy and evaluate the scheduler first on our research clusters, and then at the Ohio Super Computer Center