DYNAMICALLY MANAGING DISTRIBUTION OF DATA AND COMPUTATION ACROSS CORES FOR SEQUENTIAL PROGRAMS

摘要

Technologies are generally provided for dynamically managing execution of sequential programs in a multi-core processing environment by dynamically hosting the data for the different dynamic program phases in the local caches of different cores. This may be achieved through monitoring data access patterns of a sequential program initially executed on a single core. Based on such monitoring, data identified as being accessed by different program phases may be sent to be stored in the local caches of different cores. The computation may then be moved from core to core based on which data is being accessed, when the program changes phase. Program performance may thus be enhanced by reducing local cache miss rates, proactively reducing the possibility of thermal hotspots, as well as by utilizing otherwise idle hardware.

主权项

1. A method to dynamically manage operation of sequential programs in a multicore environment, the method comprising:
monitoring data access patterns of a first phase of execution of a sequential program initially executed on a first core of a multicore processor; determining that the sequential program has entered a second phase of execution based on the monitored data access patterns of the first phase of execution of the sequential program, wherein determining that the sequential program has entered the second phase of execution includes analyzing data access patterns during execution of the sequential program utilizing one or more of a pattern recognition technique, a machine learning technique, a support vector machine, a neural network technique, or combinations thereof; and
after determining the sequential program has entered the second phase of execution, migrating the execution of the sequential program to a second core of the multicore processor.