Optimization Of Memory Management In Virtualized Systems

In server virtualization,multiple VMs share the physical memory in one machine.Because the read and write speed of memory is usually 1 to 2 orders of magnitude higher than that of external storage devices,keeping VM's data in memory can improve its per-formance.However,research shows that the limited memory space of the server is be-coming a bottleneck that limits the density of VMs and system performance.Memory deduplication and dynamic memory adjustment for VMs are two important techniques for improving memory utilization in virtualized systems.Memory deduplication re-duces VMs' memory consumption by eliminating duplicate data.Dynamic memory adjustment improves the host's memory utilization by reclaiming memory from VMs with low requirements.In this dissertation,we analyze the shortcomings of existing methods and propose optimization solutions.The main contents are as follows:(1)Low-cost Memory Deduplication MethodExisting deduplication method(KSM)uses page comparison to find duplicate pages.However,there are many invalid comparisons in KSM.When the number of pages increases,the number of comparisons and the CPU cost increases.As a result,the efficiency of KSM decreases gradually.We propose CBF-KSM,a low-cost dedu-plication method.It filters pages before comparison by using Counting Bloom Filter algorithm,which can filter out non-duplicate pages and only allow duplicate pages to perform comparison operation.Since the number of non-duplicate pages is usually greater than that of duplicate pages in the system,CBF-KSM can greatly reduce the number of page comparisons,which reduces the CPU cost.When system memory is tight,deduplication can reduce the amount of data swapping between memory and disk,which improves application performance.The experiment shows that CBF-KSM and KSM have the same amount of memory saving when they have a same deduplication speed,and the CPU cost of CBF-KSM is only less than half of KSM.What's more,when they have the same CPU cost and system memory is tight,the application execu-tion time is shortened by up to 46%with CBF-KSM,while 30%with KSM.Compared with KSM,CBF-KSM can further improve performance by 16%.(2)Performance-aware Memory Adjustment for Multiple VMsExisting memory adjustment method(Ballooning)cannot accurately identify VMs'memory characteristics,leading to unbalanced performance among VMs.More-over,it is easy to occur excessive adjustments due to isolation,resulting in performance degradation of the VM.We propose PBPF,a dynamic memory adjustment method with VMs' performance awareness.It has three characteristics.First,it has a VM's memory state evaluation model which can identify memory characteristics of VMs and guide the memory adjustment process.Our model can reflect the difference in memory sen-sitivity of VMs,and avoid the unbalanced performance problem caused by inaccurate adjustments.Second,it is performed progressively.We mean that for a VM candidate,PBPF reclaim its memory in a fine-grained way in the early stage,then the reclaim-ing size gradually increases if there is no significant decline in performance.This can not only avoid excessive adjustments,but also ensure speed.Third,PBPF has a VM performance aware mechanism.It monitors the performance of VMs and adjusts their memory timely,which can avoid performance degradation due to excessive adjustments or sudden memory requests.Experimental results show that the performance gap be-tween the worst case and the best case of VMs is reduced from 43%to 10%,and average execution time of applications basically reaches the optimal situation with our method.