Proceedings of the NASA/ESA Conference on Adaptive Hardware and Systems (AHS), San Diego, CA, June 2011.
In this position paper, we make a case for a novel framework called MAESTRO which predictively manages system resources in shared-memory parallel computing platforms built with advanced multicore processors. In such platforms, effectively coordinating the use of asymmetric shared system resources under complex program execution scenarios becomes hard. Current resource management strategies are mostly reactive and have limited awareness of an application's resource usage and asymmetry in hardware resources. For better resource management, MAESTRO monitors the program execution environment (hardware/OS) and application behaviors, learns useful knowledge from collected information, annotates the results of the learning to relevant program and system control structures, and makes resource management decisions such as task mapping and cache partitioning in a predictive manner.