Toward Integration of Data Race Detection in DSM Systems

What are race conditions? some issues and formalizations. 29 The negligible penalties on common dsm operations, as reported in Subsection 4.1, open the way for the rst time to a full integration of data race detection with the regular operation of the dsm. Indeed, the measurements using our benchmark application suit show that the resulting overheads are acceptable. Integrating djit in the regular dsm execution makes djit a lot more signiicant than just yet another development-time debugging tool, as is the current status of data race detection systems. 28 Choi and Min 5] were the rst to propose integration of the data race detection module with the cache coherency protocol. Although they discuss hardware multiprocessors, their work can be applied in dsms. They proposed a simple algorithm, integrated with the cache manager, for on-they detection of rst data races in the fork/join model, and further extend the model with nested parallelism. In order to tag access histories they use an English-Hebrew labeling algorithm 17] and task recycling algorithm 6]. Perkovi c and Keleher 19] proposed on-they data race detection for lazy release consistency dsms. Their method instruments read and write accesses to shared variables and check for race conditions during memory synchronization. Since the system is page-based, their method suuers high slowdowns due to the process of distinguishing false from true detections. Since the technique deals with lazy release consistency systems, false detections may occur (see discussion in 1]). Similarly, pinpointing the exact location of code where the rst data race occured becomes more diicult. Richards and Larus 20] used a ne-grain dsm to get access information and track accesses between barriers. Since they support only barrier synchronization, only some of the data races can be detected online. The others are detected in a post-mortem process when execution is over, merging recorded lock and unlock times with recorded access times. Although accesses can be recorded and checked at the byte level, doing so substantially degrades performance. Hence, coarser granularity is used and the detection becomes less accurate. 6 Conclusions In this work we presented djit { an algorithm for on-they distributed detection of the rst actual data race to occur in an execution. djit makes use of an underlying ne granularity dsm which provides sequentially consistent behavior. The detection process itself is thus tuned to work in the application speciic per variable sharing granularity. This \customized" detection granularity is the …