Abstract
National Research Foundation of Korea (NRF) - Korea Government (MSIT) [NRF-2019R1A2C2009476]
Highlights
Flash memory semiconductors are increasingly being used as storage devices in smart devices, consumer electronics, and various computing platforms [1]
I/O performance is of critical importance for cloud service providers [14]: ‘‘Amazon found every 100ms of latency cost them 1% in sales
We empirically examine internal components of a flash based SSD shared by multiple virtual machines (VMs) or tasks in terms of I/O performance and proportionality
Summary
Flash memory semiconductors are increasingly being used as storage devices in smart devices, consumer electronics, and various computing platforms [1]. We empirically investigate internal components of an SSD among VMs (or tasks) sharing the SSDs as storage in terms of I/O proportionality and stability. We show the performance characteristics of SSDs in a shared storage environment and figure out the reasons of performance interference among VMs by examining the components of an SSD such as channels, DRAM buffer, and Native Command Queueing (NCQ). We find that the current Linux default I/O schedulers supporting I/O weights do not satisfy I/O proportionality requirements as requested by users and do not provide I/O stability due to performance interference between VMs or tasks. I/O proportionality here refers to the relation among the I/O throughput as observed by VMs (or tasks) competing for I/O resources in a shared storage environment.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have