A regression-based performance prediction framework for synchronous iterative algorithms on general purpose graphical processing unit clusters

摘要

Heterogeneous performance prediction models are valuable tools to accurately predict application runtime,rnallowing for efficient design space exploration and application mapping. The existing performance modelsrnrequire intricate system architecture knowledge, making the modeling task difficult. In this research, wernpropose a regression-based performance prediction framework for general purpose graphical processing unitrn(GPGPU) clusters that statistically abstracts the system architecture characteristics, enabling performancernprediction without detailed system architecture knowledge. The regression-based framework targets deterministicrnsynchronous iterative algorithms using our synchronous iterative GPGPU execution model and isrnbroken into two components: the computation component that models the GPGPU device and host computationsrnand the communication component that models the network-level communications. The computationrncomponent regression models use algorithm characteristics such as the number of floating-point operationsrnand total bytes as predictor variables and are trained using several small, instrumented executions of synchronousrniterative algorithms that include a range of floating-point operations-to-byte requirements. Thernregression models for network-level communications are developed using micro-benchmarks and employrndata transfer size and processor count as predictor variables. Our performance prediction framework achievesrnprediction accuracy over 90% compared with the actual implementations for several tested GPGPU clusterrnconfigurations. The end goal of this research is to offer the scientific computing community, an accurate andrneasy-to-use performance prediction framework that empowers users to optimally utilize the heterogeneousrnresources.