Time and Location: 2/14/2018 Wed. 2 pm, CS Conference Room 206

 

Speaker: Mr. Wei Xie

 

Title: Power-Efficient Cold Storage: IO Scheduling and Data Placement

 

Abstract: The disk power consumption is a major cost of cold storage systems. The cold storage typically has infrequent data access and high tolerance on the access time, which makes leveraging the low-power disk idle state a natural way to save power consumption. However, we observe that the disks on a cold storage system do not stay in the idle state often, even though data accesses are infrequent. We find that the IO scheduling and data placement schemes lead to such result. The disks need to wait for a deadline after an access before entering idle state. If the inter-arrival time of requests is comparable to the deadline, the disks could experience a repeatedly idle-active-idle cycle. Both the deadline and the spin-cycles reduce the effectiveness of the low-power idle state. In this work, we present an IO scheduling and a data placement scheme to increase the disk idle state times and reduce the power cycles of disks, thus reducing the power consumption as a result. The IO scheduling scheme adjusts the order of requests in the IO queue according to the power states of disks. It ensures that a span-up disk servers all (or almost all) the requests in the queue that access this disk before it enters the idle state. Compared to a FIFO scheme, it could avoid a large number of spin-cycles. In addition, due to the request queueing, disks may enter the idle state immediately after the current scheduled requests finish, instead of waiting for a deadline. For the data placement, we use a limited number of non-overlapping placement groups (of disks)  which we find to increase the effectiveness of our IO scheduling scheme significantly. We simulated the proposed IO scheduling and data placement scheme and observed a significant reduction of energy consumption with various workloads. This work could be useful for the cold storage system in public and private clouds.