Distributed Shared Memory System with Fault Detection Support

Distributed shared memory (DSM) system provides simplicity of shared-memory programming on cluster of computers. Compared to hardware approach, software DSM is a relatively low-cost, low-risk entry into the concurrent computing arena. However, fault detection becomes essential when software DSM is used to execute long-running applications. When one of the remote processor crashes or the process running on it fails, the system should provide the application with the ability to be aware of the failure and to take appropriate actions, such as cleaning up, terminating the program, and migrating the job. TreadMarks is a state-of-the-art software DSM implementation developed on standard UNIX systems. However, it does not provide fault detection support due to its usage of simple, connectionless UDP protocol. In this project, the UDP protocol used by ThreadMarks was changed to TCP protocol. TCP is connection-oriented, reliable, and of byte stream type. A fault detection mechanism using the error codes returned by the TCP socket layer was implemented. By replacing UDP with TCP, the complexity of ThreadMarks implementation can be reduced by removing redundant reliability services previously used in TreadMarks. By exploiting error codes for fault detection, we avoid any potential communication overhead because those error codes are automatically generated by TCP protocol. Our experiments show that the impact of changing UDP to TCP on the system performance depends on the scale of the application and the synchronization technique adopted by the application. Shorter-running applications and applications relying on locks for synchronization would suffer more than longer-running ones and those relying on barrier for synchronization. Our experiments also show that the fault detection using error codes returned by TCP does not incur any overhead.