BARD: Reducing Write Latency of DDR5 Memory by Exploiting Bank-Parallelism

This paper studies the impact of DRAM writes on DDR5-based system. To efficiently perform DRAM writes, modern systems buffer write requests and try to complete multiple write operations whenever the DRAM mode is switched from read to write. When the DRAM system is performing writes, it is not available to service read requests, thus increasing read latency and reducing performance. We observe that, given the presence of on-die ECC in DDR5 devices, the time to perform a write operation varies significantly: from 1x (for writes to banks of different bankgroups) to 6x (for writes to banks within the same bankgroup) to 24 x (for conflicting requests to the same bank). If we can orchestrate the write stream to favor write requests that incur lower latency, then we can reduce the stall time from DRAM writes and improve performance. However, for current systems, the write stream is dictated by the cache replacement policy, which makes eviction decisions without being aware of the variable latency of DRAM writes. The key insight of our work is to improve performance by modifying the cache replacement policy to increase bank-parallelism of DRAM writes. Our paper proposes BARD (Bank-Aware Replacement Decisions), which modifies the cache replacement policy to favor dirty lines that belong to banks without pending writes. We analyze two variants of BARD: BARD-E (Eviction-based), which changes the eviction policy to evict low-cost dirty lines, and BARD-C (Cleansing-Based), which proactively cleans low-cost dirty lines without modifying the eviction decisions. Both BARD-E and BARD-C increase bank-parallelism of DRAM writes by 30 %. As both versions of BARD have differing impact, extra misses or extra writebacks, there is no single policy that works well across all the workloads. We develop a hybrid policy (BARDH), which uses a selective combination of both eviction and writeback. Our evaluations across workloads from SPEC2017, LIGRA, STREAM, and Google server traces show that BARDH improves performance by 4.3 % on average and up-to 8.5 %. BARD requires only 8 bytes of SRAM per LLC slice.