Font Size: a A A

Optimization Design Of Distributed SSD Key-Value Storage System Based On Ensure Coding

Posted on:2020-11-27Degree:MasterType:Thesis
Country:ChinaCandidate:S XueFull Text:PDF
GTID:2428330590458330Subject:Computer system architecture
Abstract/Summary:PDF Full Text Request
Rapid development of Internet has brought about a rapid increase in the amount of data,and massive data requests require a storage system to provide rapid response and provide high availability guarantee.To improve system performance,popular data is cached in high-speed key-value storage systems to alleviate database stress.Cache failures can cause a huge increase in database load,resulting in system performance degradation or even crashes.Both Twitter's Fatcache and Facebook's McDipper are SSD-based distributed key-value caching systems that provide relatively low-cost caching for large amounts of data,but none provide reliability guarantees for the data.Erasure coding has good fault tolerance and demands small storage overhead.It is widely used in distributed storage systems to ensure data reliability.However,erasure coding needs to acquire a large amount of data when updating and repairing invalid data blocks,which brings the challenge of low response performance and high bandwidth cost to the system..EC-fatcache,a high-reliability SSD storage system based on erasure coding,is designed.The parity logging scheme is implemented in EC-fatcache to optimize the update operation.It does not update the parity in place when updating the data chunk to increase SSD life and improve update operation performance.For the problem of degraded read performance caused by parity logging scheme,the first parity optimization scheme is designed to optimize the degraded read operation,through adding a memory-based node centrally stores the first parity chunk in each strip to improve performance,reduce cost,and better adapt to read-intensive application scenarios.Experimental tests show that EC-fatcache can save more than 89% of storage resources compared to the replica solution;EC-fatcache has no increase in normal read latency compared to native systems that do not provide reliability,and write latency increases of up to 10 %;OPLG,EPLG,MPLG optimization program improved the update operation performance of EC-fatcache by more than 50%;MPLG degraded read performance increased by 40% and 24% respectively compared with OPLG and EPLG degraded read performance.
Keywords/Search Tags:Distributed storage, Key-value storage, Reliability, Ensure coding
PDF Full Text Request
Related items