An Access-Oriented Placement Strategy with Online Erasure Coding in Memory Stores
Shuang Wang, Jian Ju Luo, Yunfei Li · 2024
Objects in memory stores are becoming industry solutions for high-availability applications. These systems routinely encounter the challenges of popularity skew and server failures, which result in severe load imbalance across servers and degraded I/O performance. To solve the above problems, this paper proposes an adaptive placement scheme based on load awareness with erasure encoding for read-intensive clusters (RS-APS) and adopts the lease mechanism to optimize the data migration process. RS-APS places in-memory data and repositions them by (i) analyzing access patterns, (ii) perceiving servers’ performance, and (iii) utilizing a lease mechanism during migration. Experimental results show that compared to other placement schemes, RS-APS can increase the load balancing by more than 10X under the situation of data intensive access. The mean latency under RS-APS can be reduced by more than 2x compared to the two static schemes, i.e., the configuration-dependent placement scheme (RS-LPS) and hashing placement scheme (RS-HPS). The read latency of RS-APS with the lease mechanism can be reduced by 16.15% compared with the enhanced RS-LPS (RS-LPS+) and system throughput can be improved to 1.19x. Moreover, RS-APS exhibits obvious advantages for the scalability of storage clusters.