Deadlock detection method and system for parallel programs

Abstract

A deadlock detection method and computer system for parallel programs. A determination is made that a lock of the parallel programs is no longer used in a running procedure of the parallel programs. A node corresponding to the lock that is no longer used, and edges relating to the lock that is no longer used, are deleted from a lock graph corresponding to the running procedure of the parallel programs in order to acquire an updated lock graph. The lock graph is constructed according to a lock operation of the parallel programs. Deadlock detection is then performed on the updated lock graph.

Description

BACKGROUND[0001]1. Field[0002]The present invention relates to parallel programs, more particularly to a deadlock detection method and system for parallel programs.[0003]2. Description of the Related Art[0004]With the rapid development of computer technology, a single-core processor is gradually replaced by a multi-core processor. The multi-core processor significantly improves processing capability and computing capability of a computer, and sufficiently presents the advantage of parallel computing by integrating a plurality of execution cores into one physical processor. So called parallel computing comprises time parallel and space parallel processing, in which the time parallel is related with pipeline technology while the space parallel is related with the parallel technology performed by multiple processors. Generally, parallel computing is implemented by parallel programs in which processing of a task is separated into multiple portions (threads) and these threads can be exec...

AI Technical Summary

Benefits of technology

[0060]With the method and system of detecting deadlocks in parallel programs of the embodiments of the present invention, directed loop detection is performed based on a pruned lock graph to determine whether there is a potential deadlock, which can save a large amount of time and computing resources. Assume that a lock graph of a parallel program has N nodes and E edges, the algorithm of checking direction loop in the lock graph is divided into the following two steps: 1. find all strongly connected components in the lock graph, the complexity of the algorithm is O(N+E); 2. find all directed loops in each of the strongly connected components, the complexity of the algorithm is O(N2). Thus, the complexity of the entire algorithm is O(N2+E). It can be seen that the reduced number of nodes greatly degrades the complexity of the algorithm, and improves the efficiency of deadlock detection.

Problems solved by technology

However, process (thread) deadlock is a vital problem for parallel programs.

Generally, process (thread) deadlock will result in a breakdown of the whole system.

There are a lot of factors that will trigger process (thread) deadlock, mainly including: (1) limited system resource; (2) improper order of processes (threads) execution; and (3) unsuitable resource allocation.

If system resources are abundant and the parallel processes' requests for resources can all be met, the possibility that a deadlock occurs is low.

Otherwise, a deadlock may occur due to competing limited resources.

Further, the order and speed of advancement of process execution is different, which may also cause deadlock.

However, the resources held by T1 or T2 will not be released until the other thread releases its resources, and thus they fall in a deadlock state.

However, in practice, it is not very efficient to apply the above approach in deadlock detection since, as the program runs, more and more nodes and edges are added into the lock graph. FIG. 2 shows an example of a lock graph of parallel programs in which there are 1014 nodes and 3051 directed edges, and the operation of detecting directed loop in the lock graph is very slow, thereby consuming large amounts of time and computing resources, and greatly reducing efficiency of deadlock detection.

Images