NAND-Flash: Fast Storage or Slow Memory?

NAND-flash storage is a preferred method to accelerate I/O intensive applications because of its high throughput and fast random-read performance. However, the high cost per GB of flash, relative to magnetic disks, prevents large datacenters from moving their data entirely onto the flash media. Normally, limited amounts of flash is used “in front of” magnetic storage to speed up accesses to the persistent storage subsystem. Recently, there have been proposal that flash should alternatively be used “behind DRAM”, as a part of the volatile main memory system. In both configurations flash can be presented transparently to the application, making them competing architectural design points for where flash is to be best used in future systems. In this paper, we evaluate two state-of-the-art systems to compare the relative advantages of each approach. We present results for TPC-C, an on-line transaction processing benchmark, as database systems have been one of the largest adopters of flash technology due to the I/O intensive nature of the workloads. Our results show that when using flash for disk caching, application throughput improves by 56% and by 131% when using flash as part of a hybrid main-memory system.