Detecting anomalies in high-performance parallel programs

摘要

Message passing interface (MPI) is an effective programming technique for implementing parallel programs for distributed computation. As these applications run, a number of different types of irregularities can occur including those that result from intrusions, user misbehavior, corrupted data, deadlocks or failure of cluster components. We perform a comparison of different artificial intelligence (AI) techniques that can be used to implement a lightweight monitoring and detection system for parallel applications on a cluster of Linux workstations. We study the accuracy and performance of deterministic and stochastic algorithms when we observe the flow of function library and OS system calls of parallel programs written with MPI. We demonstrate that monitoring of MPI programs can be achieved with high accuracy and in some cases with a 0% false positive rate in real-time, and we show that the added computational load on each node is small. Finally we demonstrate that simple deterministic methods perform poorly when the program flow grows in size and variety, and that more complex methods are required.