Adam Oliner, Stanford University


A Scientific Approach to Systems Reliability

In order to build reliable systems, we must measure and analyze
existing systems. Specifically, if we hope to prevent, detect, and
predict failures, we must first understand the effects of interactions
among components. My work infers dependencies by finding anomalies
correlated in time. I use data from complex production systems
including supercomputers, autonomous driving vehicles, and software
application communities. In this talk, I will discuss recent work
aimed at detecting alerts in supercomputer system logs. Using an
algorithm that looks for high-entropy log regions (in the information
theoretic sense), we automatically detect 92% of alerts with 90%
precision on one of our data sets---this yields a maximal F1 score
that is nearly ten times that of current practice. We believe this is
both the largest-ever study of systems logs and the first reproducible
performance baseline for the task of alert detection.