The High Perfomance Scheduler Game: A Characterization of Slurm, Metrics, and the Viability of Cooperation
Tutor / Supervisor
Student
Uruchi Ticona, Wilmer Vidal
Document type
Master thesis
Date
2020
rights
Open Access
Publisher
Universitat Politècnica de Catalunya
UPCommons
Abstract
The Slurm Scheduler is a widely used tool for scheduling in High Per-
formance Computing platforms around the world. Several studies have
been conducted to nd ways to improve speci c performance metrics,
mainly from an algorithmic perspective. Scheduling has also been stud-
ied from the viewpoint of Game Theory, where models that attempt to
capture the main characteristics of the problem are developed and an-
alyzed. In this study, we have used the tools that Algorithmic Game
Theory provides to develop and study a model that captures some of the
main characteristics of the Slurm Scheduler. We developed the necessary
software to test these models. We performed a thorough data analysis pro-
cess to build a reliable data source based on real usage information. Then,
through experimentation, we analyzed how our model and its variants be-
have; furthermore, we compared these results with the results from an
existing Slurm Simulator, developed by Barcelona Supercomputing Cen-
ter members. Using these results, we calculated an approximate value for
the Price of Anarchy, and we discuss the Viability of Cooperation in the
context of the Slurm Scheduler.
