Fine Grain Algorithm Parallelization on a Hybrid Control-flow and Dataflow Processor
Nenad Korolija, Borko Furht, Veljko Milutinović
Abstract
Open-access reader
Nenad Korolija, Borko Furht, Veljko Milutinović
Abstract
Open-access reader
Abstract The execution time of a high performance computing algorithm depends on multiple factors: the algorithm scalability, the chosen hardware, the communication speed between processing elements, etc. This work is based on a hybrid processor consisting of both the control-flow and dataflow hardware, where the control-flow hardware includes manycore architecture. The algorithm decomposition into portions suitable for different architectures is presented. The results are presented by decomposing the Lattice-Boltzmann method implemented for both control-flow and dataflow hardware. The total acceleration factor of the decomposed Lattice-Boltzmann method comparing to the execution time using the control-flow and the dataflow hardware is obtained by the analysis based on the speed of the communication between both hardware equal to the speed of shared cache memories. Results indicate the advantage of using the proposed hybrid architecture and the capability of accelerating considerably even the suitable for dataflow architectures. The analysis of the problem size needed to justify the utilization of control-flow hardware is also presented. The main benefit is in accelerating algorithms for which only some algorithm portions are suitable for the dataflow hardware.
OpenAlex reports 1 citations for this work. Citation counts describe recorded attention and do not establish research quality.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
Abstract The execution time of a high performance computing algorithm depends on multiple factors: the algorithm scalability, the chosen hardware, the communication speed between processing elements, etc. This work is based on a hybrid processor consisting of both the control-flow and dataflow hardware, where the control-flow hardware includes manycore architecture. The algorithm decomposition into portions suitable for different architectures is presented. The results are presented by decomposing the Lattice-Boltzmann method implemented for both control-flow and dataflow hardware. The total acceleration factor of the decomposed Lattice-Boltzmann method comparing to the execution time using the control-flow and the dataflow hardware is obtained by the analysis based on the speed of the communication between both hardware equal to the speed of shared cache memories. Results indicate the advantage of using the proposed hybrid architecture and the capability of accelerating considerably even the suitable for dataflow architectures. The analysis of the problem size needed to justify the utilization of control-flow hardware is also presented. The main benefit is in accelerating algorithms for which only some algorithm portions are suitable for the dataflow hardware.
Key concepts: Dataflow, Dataflow architecture, Computer science, Parallel computing, Control flow, Scalability, Speedup, Data flow diagram