ALP: Alleviating CPU-Memory Data Movement Overheads in Memory-Centric Systems

Partitioning applications between near-data processing (NDP) and host CPU cores causes inter-segment data movement overhead, which is caused by moving generated one segment (e.g., instructions, functions) used in other consecutive segments. Prior works take two approaches to this problem. The first approach maps segments NDP or based on the properties of each segment, neglecting overhead. second partitions overall memory bandwidth savings, does not offload best-fitting core if they incur high movement. We show that 1) mapping its ideally can provide substantial benefits, 2) reduces benefit significantly. introduce ALP, a new programmer-transparent technique alleviate overhead systems. ALP proactively accurately transfers required key observation instructions generate stay same across different executions program. uses compiler pass identify these specialized hardware transfer their produced at runtime. evaluate wide range workloads demonstrate 54.3% 45.4% average speedup over CPU-only NDP-only executions, respectively.