Date of Original Version



Technical Report

Rights Management

All Rights Reserved

Abstract or Description

Read and write requests from a processor contend for the main memory data bus. System performance depends heavily on when read requests are serviced since they are required for an application’s forward progress whereas writes do not need to be performed immediately. However, writes eventually have to be written to memory because the storage required to buffer them on-chip is limited.

In modern high bandwidth DDR (Double Data Rate)-based memory systems write requests significantly interfere with the servicing of read requests by delaying the more critical read requests and by causing the memory bus to become idle when switching between the servicing of a write and read request. This interference significantly degrades overall system performance. We call this phenomenon write-caused interference. To reduce write-caused interference, this paper proposes a new last-level cache writeback policy, called DRAM-aware writeback. The key idea of the proposed technique is to aggressively send out writeback requests that are expected to hit in DRAM row buffers before they would normally be evicted by the last-level cache replacement policy and have the DRAM controller service as many writes as possible together. Doing so not only reduces the amount of time to service writes by improving their row buffer locality but also reduces the idle bus cycles wasted due to switching between the servicing of a write and a read request.

DRAM-aware writeback improves system performance by 7.1% and 12.8% on single and 4-core systems respectively. The performance benefits of the mechanism increases in systems with prefetching since such systems have higher contention between reads and writes in the DRAM system