A Framework for Investigating Workload Interactions on Multicore Systems

Multicore processor systems are everywhere today, targeting markets from the high-end server space to the embedded systems domain. Despite their prevalence, predicting the performance of workloads on these systems is difficult due to a lack of visibility into the various runtime interactions for shared resources. At the same time, existing support in processors for performance monitoring is primarily limited to capturing single events and thus cannot provide the depth of information required to predict dynamic workload interactions. This thesis presents a configurable and scalable multicore system framework for enabling new avenues for systems and architectural research. We have outlined and implemented a multicore platform on a Field Programmable Gate Array (FPGA), which supports the Linux Operating System (OS) and contains an integrated profiling unit, ABACUS (a Hardware Based Accelerator for Characterization of User Software), through which runtime profiling of the system can be performed. Support for systems of up to four cores has been demonstrated and the stability of the system verified through various stress tests. A system uptime of over two weeks was obtained while the system was under heavy load, and with the integration of our hardware profiler, we were able to gain insight into the operation of the system through the collection of metrics not readily possible through existing systems and simulators.