PD3: Prefetching Data with DPUs for Disaggregated Memory

We introduce PD3, a memory disaggregation solution that avoids cache misses, via prefetching, on compute servers and thus all their associated overhead. Unlike a traditional prefetcher that may pollute the cache or miss preloading opportunities due to false positives and false negatives, PD3 prevents mispredictions with network support and minimal yet critical application information. Enabling PD3 is a data processing unit or DPU, which allows (1) parsing user requests before they are processed by the compute server, (2) fetching data from remote memory on the shortest path, (3) offloading expensive RDMA and DMA operations from the host, and (4) incorporating application knowledge to faithfully predict cache misses and take actions accordingly. Designing PD3 requires reconciling DPU resource constraints and scaling requirements of cloud data systems, as well as achieving high efficiency with a myriad of performance optimizations. Our experimental results on real hardware, applications, and workloads show that with nominal compute-local memory, PD3 minimizes the performance gap between memorydisaggregated applications and their monolithic counterparts.