High-resolution analysis of parallel job workloads

David Krakov*, Dror G. Feitelson

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

3 Scopus citations


Conventional evaluations of parallel job schedulers are based on simulating the outcome of using a new scheduler on an existing workload, as recorded in a log file. In order to check the scheduler's performance under diverse conditions, crude manipulations of the whole log are used. We suggest instead to perform a high-resolution analysis of the natural variability in conditions that occurs within each log. Specifically, we use a heatmap of jobs in the log, where the X axis is the load experienced by each job, and the Y axis is the job's performance. Such heatmaps show that the conventional reporting of average performance vs. average load is highly oversimplified. Using the heatmaps, we can see the joint distribution of performance and load, and use this to characterize and understand the system performance as recorded in the different logs. The same methodology can be applied to simulation results, enabling a better appreciation of different schedulers, and better comparisons between them.

Original languageAmerican English
Title of host publicationJob Scheduling Strategies for Parallel Processing - 16th International Workshop, JSSPP 2012, Revised Selected Papers
PublisherSpringer Verlag
Number of pages18
ISBN (Print)9783642358661
StatePublished - 2013
Event16th Workshop on Job Scheduling Strategies for Parallel Processing, JSSPP 2012 - Shanghai, China
Duration: 25 May 201225 May 2012

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume7698 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


Conference16th Workshop on Job Scheduling Strategies for Parallel Processing, JSSPP 2012


Dive into the research topics of 'High-resolution analysis of parallel job workloads'. Together they form a unique fingerprint.

Cite this