Predicting Application Performance for Chip Multiprocessors

Today's computers have processors with multiple cores that allow several applications to execute simultaneously. The way resources are allocated to an application affects whether performance objectives, such as quality of service (QoS), are satisfied. To ensure objectives are met, resources must be carefully but quickly allocated in response to changing runtime conditions. \ud \ud Traditional approaches to resource allocation take place either purely online or offline. Online methods do not scale to large, multiple core systems because there are too many allocations to evaluate at runtime. Offline methods cannot handle unanticipated workloads or changes. A hybrid approach could combine the lower runtime overhead of offline approaches with the flexibility of online approaches.\ud \ud This thesis introduces AUTO, a hybrid solution to perform resource allocation. AUTO dynamically adjusts thread count, core count, and core type. It does so in accordance with a user-provided policy to meet performance objectives. AUTO's capabilities come from four prediction techniques. The first technique builds and uses models that consider CPU contention and application scalability in order to select co-running applications' thread counts. The second technique predicts applications' preferred thread-to-core mappings. The predictions are thread count independent and are translated into concrete thread-to-core mappings based on resource availability. The third technique predicts application performance under thread-to-core mappings. The final technique selects thread count and core count for applications on a system with cores of different capabilities. \ud \ud AUTO was tested in several scenarios. In each scenario, it was shown to be an effective, efficient solution to resource allocation. First, it was used to select the thread count of one or more co-running applications. Second, it was used to select application thread-to-core mappings. Third, it was used to make predictions about application performance under thread-to-core mappings. Finally, it was used to select both thread count and core type for applications on a computer with cores of different capabilities. AUTO's resource allocation and models allow for more effective and more efficient policies. By using hybrid online and offline techniques, AUTO solves the problem of allocating threads and cores to meet performance objectives.\u