Research And Optimization Of SSD Scalability In Virtualization Environment

As the development of virtualization hardware and softwaretechnology, virtualization technology has been widely used in base-serviceareas such as cloud computing. Since the purpose of virtualizationtechnology is to achieve allocation and expansion of physical resourcesdynamically, stable performance must be required while the number ofvirtual machines increasing. Scalability therefore becomes an importantmeasure of the quality of virtualization.Storage devices is one of the important physical resources, as a newkind of storage devices, SSD has a good parallel processing capabilities,however, the existing IO virtualization technology is designed undertraditional mechanical disks, so how SSD obtains scalability using parallelprocessing features in virtualization environments needs to be solvedurgently.In this paper, we measured the scalability of SSD in virtualizedenvironments, and we found two performance bottlenecks: one is the poorscalability of read request, and the other is poor performance of randomrequests.After the research of hardware of SSDs, we concluded the performancemodel for read and write requests in SSDs. We found two main reasons forthe two issued above: one is the uniform request sending in order tofacilitate the global optimization of mechanical disk drive, and the other iselevator scheduling in favor of traditional mechanical disks.Therefore we designed an optimization solution for the scalability ofSSDs in virtualization environments. Optimization program divided into thefront-end and the back-end. At the front-end, because of the different factors of overhead for read and write requests, we designed virtio_blk_mq driverto distinguish between read and write requests, improve the priority of readrequests and prevent write amplification events. At the back-end, since thereis no seek overhead of SSDs, we replace the elevator scheduling methodwith SCFQ, which uses average number of sectors accessed according tothe features of SSDs’ parallel processing.Finally, we implemented the optimization and used experiments toverify the optimization. The results of experiments also showed the effectsof each part of the optimization.