DISCS: Differentiated IO Scheduling for Cloud Storage
Random workloads bring down the effectiveness of disk utilization (bandwidth) and thus it is beneficial to isolate random workloads from sequential ones. In this project, we are building a cloud storage system that separates random workload requests from sequential ones, by directing these two types of requests to different replicas. We assume 3-way replication. Buffered writes are used to coalesce overwrites at the first two replicas. A log-structure is used at the third replica for persistency and converting random writes into sequential ones. We are building our new system based on Ceph. We have added a new replica selection algorithm (directmap) and added support in the Linux RBD device driver to detect request types and direct different types of requests to different replicas. For more details, you can check our HotCloud12 paper.
rbd driver: git-public.flux.utah.edu:/flux/git/users/xinglin/projects/fast/linux-3.2.16.git
Improve the performance of deduplication storage system
Deduplication storage systems need massive and parallel computations for hash, index lookup, block existence check, compression, and decompression operations. We propose to use GPU to accelerate these computations, which reduces overheads from these computations for read and write operations. In deduplication storage systems, files are stored in disks in nonsequential orders. However, disks are only good at sequential accesses. As a result, disks in deduplication storage systems have a significant performance degradation and increased load. For a set of Linux images we store in Venti, we can observe a significant drop(82.04%) in the read performance: the read performance drops from 34.43 MB/s to be only 6.19MB/s. We are investigating the reasons for such a huge drop and try to optimize it.
June. 2011 ~ Jan. 2012
High-performance Disk Imaging With Deduplicated Storage
In clouds and network testbeds, a disk image deployment system is needed to quickly distribute and install virtual machine images or operating system images at host devices. Previous work has shown that for these images, deduplication can save a significant amount of disk space. However, the read and write performance in deduplication storage systems is poor relative to traditional filesystem storage. In this work, we demonstrate that we can use deduplication storage systems as the backend of a high-performance image deployment system with only a negligible drop in performance by carefully pipelining to produce a balanced system.
Jan. 2011 ~ June. 2011
Refining the Utility Metric for Utility-Based Cache Partitioning
Miss rate is widely used to determine cache partitioning for multi-core systems. However, a well recognized fact in the community is that MPKI can lead to sub-optimal cache partitioning. This project is to quantify the extent of sub-optimal for MPKI based cache partitioning and proposed a simple scheme for CPI predictions.
[paper] [source code]
Linux physical memory deduplication
The main goal is to deduplicate identical pages in physical memory. We have implemented a kernel module to calculate a hash for every single physical page for both x86 and x86_64 Linux. Another kernel module is also implemented to export the content of a single specified physical page. After we found that Linux has already implemented this function in /mm/ksm.c, we stopped this project.
storage-related I/O traces:
Traces from UCSC SNIA traces
open source deduplication storage systems:
Venti ZFS opendedup
DVDStore, Microsoft Exchange Server(Loadgen), TPC-H IO profiles VDI profiles