Overcommitting Memory By Initiative Share From KVM Guests

Along with the rapid development of cloud computing industry, the computeresource management in cloud server clusters has already become the focus.Comparing to traditional server clusters, cloud computing has optimized themanagement and allocation of software and hardware resources, as well as theenhancement in hardware resource utilization and service reliability via virtualizationtechnology. However, the application of virtualization technology has encounteredsome challenges, one of which is the rationality in virtual machine memory allocation.Efficiency in virtual machine memory management is one of the key components inguaranteeing the total efficiency of cloud computing platforms. The memory resourceplays an important role as the medium between compute and storage, further memorymanagement optimizations are the key to increase the overall productivity ofvirtualization.However, since the memory management method on the current KVM platformis not able to distinguish the detailed status of guest memory pages on the host, theguest memory pages can only be detected by the visit frequency and duplication incontent on the host, which could be unreasonable, and the existing memoryovercommit strategies have different drawbacks or limits when dealing with thisproblem.A new design of memory overcommit strategy is introduced in this thesis thatmodifies the guest kernel to let it share the unused memory initiatively, and reclaimthese memory in the modified KVM module, thus eliminating the correspondinginterference to the memory scheduling algorithm on the host.This design has the capacity to utilize almost all the existing memoryovercommit mechanisms in KVM platform to simulate the operating system levelvirtualization that reserves swapping and let the host detect guest page frequencyautomatically; it simulates the loader with KSM strategy to reduce redundancy; itreclaims the unused physical memory pages on the guest like ordinary memory returnfrom applications; and it keeps pace with the guest operating system in memorybuddy priority.When comparing to other existing memory overcommit strategies, we come up with a model based on actions of a single memory page, meanwhile discussing twoimplementations and the influence from external memory fragment. Finally specifictests have been designed for the memory overcommitting strategy described above,proving its effectiveness. The macro scale experiment section has a comparisonbetween the original KVM platform and KVM with IMR to show the difference inguest system startup and dynamic memory usage scenarios, as well as the swap spaceutilization on the host.The design in this thesis is mostly made for situations with big changes inanonymous memory usage, and tests on anonymous memory usage are made in somemainstream open source software to demonstrate the basis of this thesis; when dealingwith the unused guest memory pages, the improvement is at least a thousand timesfaster than swap and KSM in microcosmic level tests.