A Dynamic Contention-aware Application Allocation Algorithm for Many-core Processor

摘要

Concurrently executing diverse independent applications on a many-core processor with hundreds of cores requests allocating application tasks to minimize communication contention and communication cost. In this paper, we propose a novel application allocation algorithm to assign applications onto a many-core processor with considering communications between tasks and contentions on network channels. Our dynamic contention-aware application allocation (DC3A) algorithm focuses on reducing both external/internal communication contentions and communication cost on network by adopting a novel edgecentric method to delicately arrange positions of tasks of an application to form a specific rectangular mapping and an efficient method to select a rectangular resource region composed of available cores to allocate the application based on the mapping. In order to evaluate DC3A, we have implemented new thread spawning/joining modules and multi-application synchronization modules in Graphite simulator. The simulation results of DC3A and peer algorithms illustrate that with the increase of communication density, DC3A can better optimize the network performance. We have observed a reduction of average packet latency (APL) at most up to 35.6%, 32.6% and 24.6% when compared with first free (FF) algorithm, nearest neighbour (NN) algorithm and contiguous neighborhood allocation (CoNA) algorithm, respectively.