Anomaly Detection System for HPC Large-Scale Parallel Programs Based on Message Passing

Abstract

The invention discloses a message-transmission based detection system for automatic monitoring of an HPC large-scale concurrent program exception and hardware-hardware cause judgment. The system overcomes the problems that a centralized detection mechanism requires the large performance cost and has poor expansibility. Through exception monitoring of a message transmission behavior, a passive heartbeat mechanism sets a message monitoring timer for a working process on each node; a heartbeat message is sent to a main control node only when the exception takes place to the message behavior; and under normal situations, heartbeats do not need to be sent. In this way, occupation of network resources is avoided; the expansibility is not limited; and state inspection is only conducted when necessary through application of a suspicious event positioning mechanism. According to the invention, the performance cost caused by execution of an MPI program can be ignored; and the system can easily expand and support judgment of software-hardware causes of running errors of large-scale concurrent application programs on a high-performance computer at running and debugging stages.

Description

technical field [0001] The invention relates to an abnormality detector applied to HPC large-scale parallel programs, and a detection system for performing software and hardware judgment on the cause of the failure of the HPC large-scale parallel programs. More specifically, it is a detection system that uses a passive heartbeat mechanism to automatically trigger abnormal alarms based on message transmission, and detects and judges the cause of abnormal software and hardware through a suspicious event location mechanism. Background technique [0002] High-performance computing (high performance computing, HPC) has a large scale, complex structure, and powerful computing power. From understanding the protein folding process to predicting short-term and long-term climate patterns, massively parallel HPC simulations have become the tool of choice. These applications can Run detailed numerical simulations to model the real world and enable breakthroughs in science and engineerin...

AI Technical Summary

Problems solved by technology

[0010] In order to determine the root causes of HPC large-scale parallel program execution failures, the purpose of the present invention is to provide a detection system based on message passing-based automatic monitoring of program exceptions and discrimination of software and hardware causes. message transmission, using the passive heartbeat mechanism, to realize the abnormal automatic alarm and trigger the location of suspicious events during the execution of HPC large-scale parallel programs; , high overhead, and poor scalability; on the other hand, it realizes the automatic alarm and detection of abnormalities in the execution process of HPC large-scale parallel programs, accurately locates hardware faults, and provides the most likely candidates for the root cause of software errors. item

Images