Triple A: application-aware aggregation for leveraging internal parallelism of key-value SSD

Abstract

Key-Value (KV) store has been used widely because of its flexible and straightforward interface. In this trend, Key-Value SSD (KV-SSD) has been proposed to minimize host SW involvement by offloading KV functionality onto the storage device. It shows better application-level write performance with fewer resource requirements than the case of using Block-SSD. However, it represents worse read performance. To solve this problem, we introduce a new opportunity in KV-SSD about SSD internal parallelism not issued. Although particular value sizes exist in which the KV-SSD fully leverages its parallelism, this characteristic has not been used as a crucial design point so far. In this paper, we propose "Application-Aware Aggregation" approach and system Triple A to leverage the internal parallelism of KV-SSD and improve the performance. Our approach's primary idea is to use properties about sequentiality and spatial locality, which can be found in several applications, in order to reduce aggregation overhead and increase read efficiency. With two sample applications, Triple A achieves x500 throughput & x0.55 completion time compared to basic KV-SSD. It also achieves x1.78 throughput & x0.57 completion time than LSM-based-tree KV stores while using less than or equal to the CPU resources.