|Institution:||University of Minnesota|
|Keywords:||Energy saving; Hybrid cache; RAID; RAM; Electrical engineering|
|Full text PDF:||http://hdl.handle.net/11299/170837|
In this paper, we propose a method of using hybrid cache to improve the cache hit rate of popularity-based caching management scheme for Redundant Array of Inexpensive Disks(RAID) architecture. The hybrid cache we proposed consists of Solid State Drive(SSD) and battery-backed Dynamic Random Access Memory(DRAM). The main purpose of using SSDs as cache for RAID is for energy saving in this thesis, however, SSDs have few special features: 1. SSDs do not support in-place updates, thus a erase operation has to be performed before a page gets updated. 2.SSDs have relatively low endurance, so frequent erasures and write operations can wear out SSD units and reduce the life expectation of SSDs. Dongkyu Lee et al  proposed a popularity-based management scheme(PBM) to reduce write operations on SSDs cache, this scheme works well under web-sever-based environment where a small number of hot files are accessed frequently. However, because PBM calculates the popularity of files and caches popular files into SSD cache only at the end of each time period, it is not able to exploit the temporal locality hidden in the I/O requests, which leads to a reduced cache hit rate. In this paper, we propose a method of using a SSD-Battery Backed DRAM hybrid cache to reach both a high cache hit rate and low write size count on SSD. DRAMs have very good endurance compared to SSD, and DRAMs also support in-place update. So we can take advantage of high endurance and in-place update of DRAMs to explore the temporal locality of user requests. Our SSD-DRAM Hybrid Management Scheme(SDHM) will implement PBM in SSD and simple replacement caching algorithm like.Least Recently Used(LRU), Least Frequently Used(LRU) and Random Replacement in DRAM. While popular files in SSD can service the frequent read requests on popular files, DRAM with simple replacement algorithm can help to exploit the temporal locality. Also, Considering the volatility of DRAM, we use battery-backed DRAM in our SDHM in case of sudden power breakdown. We simulated both SDHM and PBM along with few other caching schemes using real traces, the results show that, compared to PBM, our SDHM maintains both a high cache hit rate and a low SSD write count.