Talk Title: LSVD: Elastic Storage on the Gateway

Talk Abstract: Storage systems can be measured in two dimensions — capacity and performance. A single storage device or storage-equipped server typically provides these in a fixed ratio, with improvements in one resulting in either an increase in capital cost or a decrease in the other. Yet demands for storage capacity and performance vary widely, both across workloads and over time.
Today’s disaggregated storage systems decouple storage capacity from CPU resources, allowing client computation and (static) storage requirements to be met without stranding resources of either class. Yet their designs typically couple storage capacity and performance, in a ratio which must be chosen at the time of initial deployment. The result is either stranded performance (i.e.high per-TB cost) or stranded capacity under high-performance workloads.
LSVD decouples the performance-related components of the storage system — caches and per-I/O CPU usage — from its capacity tier, allowing performance to be provisioned independently of capacity and even elastically as workloads change over time. To demonstrate this, our evaluation shows SSD-class performance over HDD backends, using a scalable intermediate caching tier to provide scalable performance.
Virtual machines running on remote virtual disks are the backbone of many cloud services, and are thus a key component of an Open Cloud. Yet today’s virtual disks suffer from high write overhead due to replication, metadata and consistency protocols, yielding cheap storage speeds at high-performance storage prices. In contrast, gateway LSVD gives high write performance without paying a “performance tax” on capacity. We are working with IBM to deliver low-overhead high-performance Ceph-based volume storage based on LSVD, with a planned experimental rollout in NERC late this year.

Bios: One of the MOC founders, Peter’s research focuses on data storage in operating systems and the cloud, while his teaching seeks to help students understand computer systems “under the hood”. A professor at Northeastern since 2008, his prior career spans the years and alphabet from Apple to VMware, with a bevy of startups in between.

Isaac is a PhD Student at NEU with a focus on storage and operating systems. Before starting his PhD, he worked on XDP and BPF technologies at Intel, and was a system administrator for multiple large-scale clusters. His current work focuses on scalable storage and on improving the Linux kernel for storage applications.