Write Performance Optimization On TRPP3

Data loss will cause massive loss to person and company, storage system should bereliable enough to protect user’s data. RAID(Redundant Array of Independent Disks) canprotect data in spatial dimension and CDP(Continuous Data Protection) can protect data intime dimension to prevent data damaged by virus attacks, user errors. TRAP3combinedRAID5and CDP to provide a comprehensive data protection solution. TRAP3has to storeold data of each block to backup device when writing new data to the block, so TRAP3choose Read-Modify-Write all the time to process write request for the sake of using theold data read by Read-Modify-Write. TRAP3has a lower write performance thanRAID5because of Reconstruction-Write ability loss.To improve TRAP3’s write performance, WPOT3(Write Performance Optimizationon TRAP3) is proposed and implemented. WPOT3has different stripe initial process,write procedure and data recovery process with TRAP3. WPOT3write0to each block atstripe synchronization stage. There are two purpose to write each block0. One formaintaining the XOR relationship of a stripe’s data block. The other purpose is to allowWPOT3to write new data of a block to backup device rather than backup old data. Thanksto the modification of write procedure,WPOT3can choose Read-Modify-Write orReconstruction-Write as appropriate. WPOT3also need to alter the data recovery processbecause of the change on write procedure.WPOT3is designed and implemented by altering Linux RAID5module.Comprehensive tests are done to RAID5,TRAP3,WPOT3including sequential write,random write, sequential read, random read,sequential write of variable write proportion,and random write of variable write proportion. Test results shows that the writeperformance of WPOT3is obvious higher than TRAP3,lower but close to RAID5.WPOT3is especially suitable for processing intensive workload of sequential write andintensive workload of big data random write. Compared to TRAP3, WPOT3cansignificantly improve the write performance on these two workload.