MeLoPPR: Software/Hardware Co-design for Memory-efficient Low-latency Personalized PageRank
Hosted in Virtual Platform
Embedded System Design Methodologies
DescriptionPersonalized PageRank (PPR) is a widely used algorithm in real-world applications that requires low latency computation. Most existing works focus on algorithmic optimization for improving accuracy or global graph processing on large-scale systems for improving throughput. While minimizing local PPR latency with a tight memory budget remains unexplored, we propose a hardware-friendly multi-stage PPR with much less on-chip memory requirement through stage and linear decomposition. We also propose a hybrid CPU and FPGA accelerator that greatly shortens the latency. We evaluate our framework on both CPU and Xilinx ZCU102 FPGA, and demonstrate remarkable speedups with greatly reduced memory.