Management And Optimization Of Heterogeneous Storage Under Virtualized Computing Environment

With the scale increasing and the usage diversifying,the computer storage system has shown the characteristic of heterogeneity.It often requires software to have enough awareness of the underlying architecture and global resource usages to take the full advantage of the heterogeneous storage system,which is hard for the virtualized environment.As one of the most important platforms of the cloud service,virtualized environment is in desperate need of being improved on performance and utilization when running on top of the heterogeneous storage system.This paper discusses the management and optimization of heterogeneous storage resources from three aspects,Non-Uniform Memory Access architecture,Non-Uniform I/O Access architecture as well as persistent memory.In the NUMA/NUIOA architecture,different processor nodes are directly linked with different memory controllers and I/O controllers.While this design enables the parallel processing of multiple memory access requests and device access requests,it also forces a request to be redirected between nodes via interconnects if the request issuer does not reside on the same node of the request receiver.The persistent memory presents characteristics other than the persistence of on-chip data,such as asymmetric read/write performance and limited durability,which is determined by the implementation of persistent memory.To summarize,the performance and utilization of the heterogeneous storage in the computer system are highly relevant to the policy of resource allocation and resource management.However,the semantic gaps brought in by the virtualization makes it hard for both the resource users and resource managers to manage and allocate hardware resources according to the workload and the global resource usage information.This paper includes following works.Propose a system to characterize the in-VM workload with hardware performance monitor,and scheduling policy based on the characteristics of in-VM workload.According to the evaluation,the system can increase the local memory access ratio / decrease the last level cache miss rate,which leads to the improvement of the memory access performance.Design a set of interfaces between the virtual machine and hypervisor to pass the in-VM I/O access semantics to the hypervisor.Propose scheduling policy to reduce the ratio of remote I/O accesses and improve the overall performance under NUIOA architecture.Propose observations on the necessity of virtualizing persistent memory.Design frameworks to manage the persistent memory for virtualized environment,which provides an abstraction of virtualized persistent memory to VM and characterize the pattern of persistent memory accessing.Propose memory allocation strategies based on persistent memory access patterns.According to the evaluation,the system can increase the utilization of persistent memory and reduce the amount of persistent memory required for the same workload,while maintaining the non-volatility and performance of persistent memory.