大容量DRAM的刷新开销问题及优化技术综述

动态随机存储器(DRAM)具有速度快、密度高、成本低的优势，被广泛应用于计算机的主存.DRAM采用电容作为存储单元，电容电荷的多少表示数字“0”或“1”.由于存在漏电现象，电容里的电荷会缓慢流失，造成数据丢失.为保证数据正确性，DRAM采用周期性的刷新操作，在数据丢失前，把数据读出然后重新写入存储单元.刷新操作会阻碍正常访存的执行，造成性能上的开销；同时刷新操作会消耗额外的功耗，带来功耗上的开销.刷新的开销与DRAM密度相关：在过去，当DRAM密度较小时，需要刷新的存储单元数较少，刷新开销很小，并未引起关注；但是，随着摩尔定律的发展，DRAM密度越来越大，目前已发展到千兆比特级别，其刷新周期并没有改善，单位时间内需要刷新的存储单元数越来越多，从而使刷新带来的性能和功耗开销越来越严重.刷新问题目前得到了工业界和学术界的广泛关注.首先介绍了目前DRAM的刷新方式和开销，以及工业界已经实现的一些改进；然后把工业界和学术界提出的众多优化方法分为“减轻刷新操作对访存的阻塞”和“减少不必要的刷新操作”两大类，分别进行了分析和总结；最后给出了关于智能刷新管理的总结和展望.

Problems and Optimizations of Refresh for Large-Capacity DRAM

DRAM (dynamic random access memory) is widely used as main memory of computer system, which is of fast speed, high density and low cost. DRAM uses capacitors as basic storage cells, and uses the amount of charges to represent digital “0” and “1”. However, the capacitor charges leak over time, causing data lost. To maintain data integrity, DRAM periodically refreshes all cells-read data out before lost and rewrite into cells. Refresh operations block normal memory requests, causing performance overhead; refresh operations also consume extra power, causing energy overhead. The refresh overheads are related to DRAM density. In the past, DRAM density was relative small, and the amount of cells needing to be refreshed was not that large, so the overheads gain little attention. But as the evolving of Moore’s Law, DRAM density grows to Gigabits today, and more cells need to be refreshed during the same period, exacerbating the performance and energy overheads. The problem of refresh has now been an important concern for both industry and academia. In this paper, we first introduce how refreshes are performed, its overheads, and some improvements from industry; then we classify the many improvements from industry and academia into two categories-reducing the blocking of memory requests, and reducing the unnecessary refreshes-and give our analysis and summaries, respectively; finally, we conclude the research work and point out the possible research directions.