Sievestore: a highly-selective, ensemble-level disk cache for cost-performance
Abstract
The storage layer is a significant factor in the capital cost, energy cost and performance of servers and data-centers. Emerging high-performance, low-energy, non-volatile, solid-state storage media show enormous potential to solve the storage performance and energy problems. However, the high cost-per-byte of solid-state media has hindered wide-spread adoption in servers. Thus we propose a new, cost-effective architecture - SieveStore - that enables the use of solid-state media to significantly filter access to storage ensembles. This work makes three key contributions. First, we make the case for highly-selective, storage-ensemble-level disk-block caching based on our observations that block accesses of storage ensembles exhibit extremely skewed popularity, and though the degree of skew for the ensemble as a whole remains invariant, the popular block set is dynamic across storage volumes of both the same and different servers, and in time for the same server. Second, we show that selective cache allocation - sieving - fundamentally enables efficient ensemble-level disk-caching. Third, we propose two variants of SieveStore. The first variant SieveStore-D) selectively batch-allocates cache space at discrete time-intervals. The second variant (SieveStore-C) allocates cache space continuously via a hysteresis-based, lazy allocation policy. Based on traces from a storage ensemble representing the block accesses of 13-servers, over a period of one week, we find that sieving and ensemble-level caching each contribute to SieveStore's cost-effectiveness. Both variants of SieveStore achieve significantly higher hit ratios than unsieved ensemble-level disk-caching (35%-50% more, on average) while using only 1/7th the number of SSD drives. Further, ensemble-level caching is a strictly superior cost-performance point compared to per-server caching.
Degree
M.S.E.C.E.
Advisors
Thottethodi, Purdue University.
Subject Area
Computer Engineering|Computer science
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.