Abstract
Heterogeneous computer architecture is becoming the mainstream of computer systems. In the landscape of heterogeneous computer architecture, General-Purpose Computing on Graphics Processing Units (GPGPUs) is indispensable for supporting ultra high density computing. As the complexity of GPGPUs increases, designing efficient GPGPUs requires good tools to support. Among them, execution-driven simulators are often used to evaluate the architectural designs of GPGPUs. Execution-driven simulators can provide quite accurate and comprehensive performance data, but they often require very long simulation time, which slows down the process of design space exploitation. On the other hand, trace-driven simulators simulate only the speci c components that are of interest, e.g. Network-on-Chip (NOC) or cache hierarchy, and rely on execution traces to mimic the operations of other components, e.g. processor cores. As a result, trace-driven simulators are fast and suitable for design space exploitation. However, traces are the execution results of the trace-generating machines, not target machine. Thus, trace-driven simulators often produce performance data that have large error margins. A recent trend in trace-driven simulation is to use the causality relationships among the trace events to adjust the event timing, instead of using the absolute event time from the trace-generating machines. In this thesis, we apply the con- cept of causality-aware trace-driven simulation to the evaluation of the NOC of GPGPUs. We take a widely used execution-driven GPGPU simulator, Multi2Sim, and study how to extract causality information from its execution trace. One difficulty in determining the causality relationships of NOC events for GPGPUs is the latency hiding mechanism, which allows multiple memory access requests outstanding at the same time. We discuss how to leverage the memory fence instructions of GPGPUs to identify the causality relations. The extracted causality traces are then fed into a well-known trace-driven NOC simulator, Garnet, which is modi ed to be causality-aware. Our evaluation results show that the causality-aware Garnet can match the performance trend obtained from the execution-driven simulator Multi2Sim, while the original Garnet cannot.