A transparent monitoring tool for shared-memory multiprocessors

Monitoring and debugging of parallel programs is complicated by race conditions, which can cause software monitoring to alter program behavior. To avoid these unwanted modifications of program execution, the authors present a flexible scheme for transparently monitoring parallel programs in a shared-memory environment. To achieve transparency, the monitor observes causal relations between events in different threads of execution, and intervenes when an impending event would change the order of occurrence of causally related events, as compared to unmonitored execution of the same program. Constructs used to support this monitoring scheme are developed, including mechanisms to deal with unsynchronized and coarse grained clocks. The monitoring scheme requires the instrumentation of every shared-memory access. To measure the overhead created by this intrusion, a prototype monitor has been implemented. Preliminary performance results produced by the prototype are presented and discussed