Loading Events

« All Events

  • This event has passed.

ECE PhD Dissertation Defense: Mohammad Hossein Hajkazemi

May 25, 2021 @ 10:00 am - 11:00 am

PhD Dissertation Defense: High-performance Translation Layers for Cloud Immutable Storage

Mohammad Hossein Hajkazemi

Location: Zoom Link

Abstract: Most storage interfaces support in-place updates: blocks can be rewritten, files can be modified at byte granularity, fields may be updated in database table rows. Yet internally these layers often rely on out-of-place (immutable) writes. In some cases, this may be necessary to use media, such as flash, SMR (shingled magnetic recording) and IMR (interlaced magnetic recording) disk, which do not allow overwrites. In others, it is used to simplify the implementation of transactions and/or crash consistency, in the form of journaling, write-ahead logging, shadow paging, etc.

In a storage system, translation layers perform out-of-place writes, and they are implemented in different layers of storage stack from the file system to the storage device firmware depending on the application. In this dissertation I focus on translation layers for cloud immutable storage technologies to improve the cloud I/O performance. As a part of my thesis, I focus on translation layers for state-of-the-art immutable storage media such as SMR and IMR used in cloud environments, proposing several novel algorithms to improve their efficiency. I also introduce FSTL, a framework to design and implement SMR translation layer. Finally, I describe Collage, a virtual disk I developed over S3 (could be implemented over a similar object storage) using a translation layer which performs large, sequential, out-of-place writes for high performance. It optionally uses fast local storage for write logging and as a write-back cache, guaranteeing prefix consistency under all failure conditions and recovery of all acknowledged writes if the local cache is not lost. Collage supports snapshots and cloned volumes, performs well over erasure-coded storage, and allows consistent asynchronous volume replication over geographic distances. I show that Collage can achieve massive performance improvements (e.g., over 100x for microbenchmarks and 10x for macro-benchmarks) over CEPH RBD, a popular open-source scale-out virtual disk implementation.

Details

Date:
May 25, 2021
Time:
10:00 am - 11:00 am
Website:
https://northeastern.zoom.us/j/95483290933#success

Other

Department
Electrical and Computer Engineering
Topics
MS/PhD Thesis Defense