Measurement and Modeling of EARTH-MANNA Multithreaded Architecture

In this paper, we develop and apply an analytical model to predict the performance of McGill's EARTH-MANNA multithreaded multiprocessor system. The performance model is evolved from a closed queuing network model for multithreaded architectures reported in our earlier work [17]. In this work, we extend the original model to account for the complications due to realistic subsystem interactions and multithreaded workload in the EARTH-MANNA system. Our solution technique uses mean value analysis. We use simple heuristics to account for the subsystem interactions (simultaneous resource possession) and multithreaded workload. Inputs to the analytical model are architectural parameters derived from the EARTH-MANNA system measurements, and program workload parameters like number of threads, and their runlengths, supplied by the users. Outputs of the model include performance predictions such as processor utilization and network latency for remote accesses with split-phase transactions. We characterize the variation of these performance measures in terms of program workload and architecture characteristics. Our analytical model has been validated through runtime measurements on the EARTHMANNA system. A set of accurate software-controlled instrumentation tools has been developed to measure the network latency and processor utilization during actual multithreaded program execution. The runtime measurements from EARTH-MANNA system on synthetic benchmarks match well with the analytical model (within 10% of the model predictions). The model predictions also conform with the measurements on real programs (in [16]). Our results demonstrate the tradeo s of realistic costs of multithreading on the performance for ne-grain parallel program workload. The model is useful for both users (compilers) and system architects of multithreaded architectures and systems.