Adaptively Separating Read And Write Upon SSDs Of Hybrid Storage Nodes

Nowadays,large cloud providers usually deploy hybrid storage nodes consisting of Solid State Drives(SSDs)and Hard Disk Drives(HDDs)in their cloud datacenter,to reduce the overall cost on the premise of ensuring the service quality.These hybrid storage servers generally employ SSDs as primary storage serving all requests from front-end applications,while using HDDs as the secondary storage to provide sufficient storage capacity at low cost.This storage mode is called SSD First Level(SFL)mode.By analyzing real production workloads from Pangu,a large-scale distributed storage system,we find that,under mid/high intensive reads,even a small number of writes dramatically increase read-latency,especially for tail latency,and they are caused by write-induced-blocking and write-induced-garbage-collection(GC).To address the problem,we present a runtime IO scheduling approach called SeRW that strategically steers SSD-writes to idle HDDs at runtime.Specifically,SeRW monitors the workload intensity and all request queues of SSDs and HDDs at runtime.When SSD-reads are under mid/high intensive,if the SSD queue length is larger than the threshold,SeRW attempts to redirect all SSD-writes to idle HDDs,else it only redirects the SSD-writes with large size to idle HDDs.SeRW decreases the write-blocking read latency on SSDs and reduces the amount of data written into SSDs,thus decreasing the GC frequency.We evaluate SeRW through replaying the production workload.The experimental result shows that SeRW decreases the average,99^th,99.9^th and 99.99^th-percentile latencies of SSD-reads by up to 2.07x,1.48x,4.29x,and 4.24x,respectively,without sacrificing the write performance,while reducing the amount of data written to SSDs by up to 37.5%.