DRAMHiT: A Hash Table Architected for the Speed of DRAM
Reading group: Teodor Cvijovic presented "DRAMHiT: A Hash Table Architected for the Speed of DRAM" (Eurosys'23) at 1D19 the 2/2/2024 at 10h00.
Abstract
Despite decades of innovation, existing hash tables fail to achieve peak performance on modern hardware. Built around a relatively simple computation, i.e., a hash function, which in most cases takes only a handful of CPU cycles, hash tables should only be limited by the throughput of the memory subsystem. Unfortunately, due to the inherently random memory access pattern and the contention across multiple threads, existing hash tables spend most of their time waiting for the memory subsystem to serve cache misses and coherence requests.
DRAMHiT is a new hash table designed to work at the speed of DRAM. Architecting for performance, we embrace the fact that modern machines are distributed systems— while the latency of communication between the cores is much lower than in a traditional network, it is still dominant for the hash table workload. We design DRAMHiT to apply a range of optimizations typical for a distributed system: asynchronous interface, fully-prefetched access, batching with out-of-order completion, and partitioned design with a low-overhead, scalable delegation scheme. DRAMHiT never touches unprefetched memory and minimizes the penalty of coherence requests and atomic instructions. These optimizations allow DRAMHiT to operate close to the speed of DRAM. On uniform key distributions, DRAMHiT achieves 973Mops for reads and 792Mops for writes on 64-thread Intel servers and 1192Mops and 1052Mops on 128-thread AMD machines; hence, outperforming existing lock-free designs by nearly a factor of two.