Building JACK: Developing Metrics for Use in Multi-Objective Optimal Data Allocation Strategies

Published as Storage Systems Research Center Technical Report UCSC-SSRC-14-01.

Abstract

Current file systems optimize for a single objective, often to the detriment of other objectives. This can result in an unbalanced system that does not reflect the needs of the system or its users. We suggest addressing the data allocation problem as a multi-objective optimization problem, using our general data placement framework, JACK. This provides us with a means of finding an optimal data allocation, tailored to the requirements of individual systems, requiring at most the specification of the relative importance of competing performance goals. For such an approach to be feasible, we need to define meaningful, yet observable, metrics for such a multi- objective optimization problem. We introduce the objectives we intend to optimize for and example metrics which can be used to gauge them. We provide a detailed description of how each metric was developed, using a data set provided by Los Alamos National Laboratory for evaluation. We also explain how the metrics enable the development of JACK, with the intent that others can make use of our metrics and develop their own metrics for additional performance goals.

Publication date:
January 2014

Authors:
Christina Strong
Ahmed Amer
Darrell D. E. Long

Projects:
Dynamic Non-Hierarchical File Systems

Available media

Full paper text: PDF

Bibtex entry

@techreport{strong-techreport14,
  author       = {Christina Strong and Ahmed Amer and Darrell D. E. Long},
  title        = {Building {JACK}: Developing Metrics for Use in Multi-Objective Optimal Data Allocation Strategies},
  institution  = {University of California, Santa Cruz},
  number       = {UCSC-SSRC-14-01},
  month        = jan,
  year         = {2014},
}
Last modified 28 May 2019