Dynamic Heterogeneous scheduling of GPU-CPU in Distributed Environment

Abstract

The technology has been growing in the field of computation so fast that the operation which takes days in old time now can be completed within few seconds. The basic need of such computing is fastest processing time of any operation by the system. The performance of any computing device is dependent on processor, memory and hardware/ software behavior. Central Processing Unit (CPU) is known to be the brain of computer. In any case, progressively, that brain can be raised by additional piece of Personal Computer with the GPU (Graphics Processing Unit), which is alluded to as the aforementioned spirit. The fusion of a CPU with a GPU can convey the ideal estimation of framework cost, execution and power. It tends to be expressed that GPUs vary from CPUs as GPU is improved for throughput rather than idleness which can work quicker and more expense effectively than CPUs. GPUs are equipped for taking a lot of information and playing out a similar task again and again rapidly, in contrast to CPU, which will in general skip activities everywhere. Distributed processing comprises of utilizations successively over a stage which have more than one computational device with various designs, for example, a manycore GPU and a multi-core CPU. By and large, the kernel performs well on the GPU as they are enhanced for a GPU's exceedingly with parallel engineering and GPU regularly provide higher pinnacle throughput per unit of time. The exploration says that GPU is definitely more superior to CPU because of its parallel design as it is made out of hundreds of cores which can deal with a huge number of threads when contrasted with CPU. Here, we will demonstrate this fact that GPU is growing its importance in High-Performance Computing (HPC) era. In this paper, we will take few applications with the historical information about the runtime of particular application taken on CPU and GPU. This historical information created from benchmarks will let help us to decide whether the tasks are GPU bound or CPU bound and schedule them accordingly to reduce the waiting time of other applications. Our approach immensely takes dynamic decision to schedule the tasks. Earlier approaches are not as impactful as our approach because here the greedy decision taken to reduce overall execution time and improve processor utilization.