2021/05/21 by Houssam-Eddine Zahaf, Zahaf, Houssam-Eddine, Ignacio Sanudo Olmedo +7
Computer Science · #Distributed #Distributed and Parallel Computing Systems #FOS: Computer and information sciences #FOS: Electrical engineering #Interconnection Networks and Systems #Parallel #Parallel Computing and Optimization Techniques #Systems and Control (eess.SY) #and Cluster Computing (cs.DC) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2105.10312
openalex publication_date 2021/05/21 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
In order to satisfy timing constraints, modern real-time applications require\nmassively parallel accelerators such as General Purpose Graphic Processing\nUnits (GPGPUs). Generation after generation, the number of computing clusters\nmade available in novel GPU architectures is steadily increasing, hence,\ninvestigating suitable scheduling approaches is now mandatory. Such scheduling\napproaches are related to mapping different and concurrent compute kernels\nwithin the GPU computing clusters, hence grouping GPU computing clusters into\nschedulable partitions. In this paper we propose novel techniques to define GPU\npartitions; this allows us to define suitable task-to-partition allocation\nmechanisms in which tasks are GPU compute kernels featuring different timing\nrequirements. Such mechanisms will take into account the interference that GPU\nkernels experience when running in overlapping time windows. Hence, an\neffective and simple way to quantify the magnitude of such interference is also\npresented. We demonstrate the efficiency of the proposed approaches against the\nclassical techniques that considered the GPU as a single, non-partitionable\nresource.\n