The Creation Of Redundant Memory Space For High Availability

The scale of cloud platforms is becoming larger and larger, and theiravailabilities are becoming more critical. Investigation shows that cloudservice provider will lose millions of dollars per hour if a cloud platform isout of service. But recent research indicates that memory errors will occurin about32%of the machines per year, and1.3%of them will crash due tothose errors. That means memory errors should be regarded as commonsituations in large clusters, and need to be solved as soon as possible.The traditional solutions for memory availability can be divided intotwo categories: hardware approaches and software approaches. Thehardware approaches include mirroring memory and ECC memory; bothof them need special and expensive devices to support. The softwareapproaches are mainly based on backup data on application level or basedon backup whole system via virtualization technology; but theperformance is relatively low.This dissertation aims to enhance memory availability using softwaremirroring memory. It creates the redundant space for others to store thereplicated data. The system is implemented in two separated approaches.The first is an efficient, lightweight, operation system (OS) based approach,which is called Memv6; the second is a virtual machine monitor (VMM)based, OS independent approach, which is called Memvisor.By modifying memory management module in the OS, Memv6willcreate the redundant page table while the native page table is created. Thisdissertation also analyzes the memory layout in OS, and introduces themodification of it to ensure that there is no overlap between native addressand redundant address. Memvisor is implemented in VMM level. By modifying the processof page table creation with direct page table technology, or modifying theprocess of page table translation with shadow page table technology,Memvisor can create the related redundant space. This modification istransparent to the OS, which enhances the usability of the system.Furthermore, high-available virtual machines and native virtual machinescan be run in one physical node simultaneously, which is more flexiblethan hardware high-availability solutions.The evaluation shows that all backup spaces have been createdsuccessfully. It only causes less than1ms to execute this process for200KB space. The overhead of xv6commands is5%~8%, and theoverhead of a tiny web server, thttpd, is less than5%. Compared withtraditional software high-available solutions, Memv6and Memvisor onlycause about5%to10%overhead of them.