Bilkent University
Department of Computer Engineering


Domain Specific Processor for Graph Applications Using Scratchpad Memory


Gülce Pulat
MS Student
Computer Engineering Department
Bilkent University

Working on large graphs require main memory to be accessed very frequently and in a scattered fashion, which results in poor data locality and memory bottleneck. To tackle this bottleneck we introduce a customized processor that has a scratchpad memory (SPM) utilized according to the access patterns of graph applications. We modified an open source RISC-V core and improved it with SPM and SPM controller units. To obtain machine codes for our core and to show how our SPM can be used in a graph algorithm, we extended RISC-V GCC compiler with SPM related instructions and we developed a new BFS code. Simulation results give hit rates around 95%, showing that the graph processor can indeed increase performance.


DATE: 16 December 2019, Monday @ 16:10