Bigger, Longer, Fewer: What do cluster jobs look like outside Google?Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-17-104, October 2017.
George Amvrosiadis, Jun Woo Park, Gregory R. Ganger, Garth A. Gibson, Elisabeth Baseman*,
Carnegie Mellon University
* Los Alamos National Laboratory
In the last 5 years, a set of job scheduler logs released by Google has been used in more than 400 publications as the token cloud workload. While this is an invaluable trace, we think it is crucial that researchers evaluate their work under other workloads as well, to ensure the generality of their techniques. To aid them in this process, we analyze three new traces consisting of job scheduler logs from one private and two HPC clusters. We further release the two HPC traces, which we expect to be of interest to the community due to their unique characteristics. The new traces represent clusters 0.3-3 times the size of the Google cluster in terms of CPU cores, and cover a 3-60 times longer time span.
This paper presents an analysis of the differences and similarities between all aforementioned traces. We discuss a variety of aspects: job characteristics, workload heterogeneity, resource utilization, and failure rates. More importantly, we review assumptions from the literature that were originally derived from the Google trace, and verify whether they hold true when the new traces are considered. For those assumptions that are violated, we examine affected work from the literature. Finally, we demonstrate the importance of dataset plurality in job scheduling research by evaluating the performance of JVuPredict, the job runtime estimate module of the TetriSched scheduler, using all four traces.
KEYWORDS: cloud, clusters, study, workload, scheduling, traces, JamaisVu, heterogeneity, utilization, failure rates
FULL TR: pdf