Performance Analysis And Optimization For Virtual Machine Level Fault Tolerance

With the development of cloud computing,more and more application services are applied on the cloud platform.In order to achieve high availability,cloud providers use replication technology for fault tolerance.Virtualization can encapsulate the application into a virtual machine,and then achieve high availability by periodically backing up the whole VM states to the replication.Currently,VM fault tolerance mechanism has become an important research topic.However,there are still some problems in the existing VM fault tolerance mechanism.First,in order to maintain external consistency,the replication system suffers from the roundtrip delay of network packets and the checkpoint overhead.Second,due to the unnecessary consumption of CPU resources,the backup VM may lead to the double scheduling problem in the virtualized environment.This paper proposes two key technologies to address these problems: 1)packet early releasing technology,2)active/semi-active VM replication model.The former analyzes the connection between packets and memory areas,and early releases the packets if they are clean.The latter one mitigates the double scheduling problem by dynamically reducing the number of active v CPUs(virtual CPUs)in the backup VM and converting the active/active replication model into an active/semi-active one.The main contribution of this paper consists of the following parts:1.This paper delves into the existing fault tolerance mechanisms and finds out several problems.In order to protect the integrity of the data,existing replication system adopts packets comparison and checkpoint to maintain external consistency.These operations result in large performance gaps between the replication systems and one single VM.In this paper,we propose the packets early releasing technology,and it improves the performance of the replication system in some scenarios;2.We study the active/active replication model of the original system and find out that the intention of this design is to compare packets and reduce the overhead caused by the checkpoint.However,this model consumes a large amount of CPU scheduling resources on the backup machine and introduces the double scheduling problem in a virtualized environment.To address this,we propose the active/semi-active replication model.Without affecting the overall performance of the system,this model dynamically reduces the number of active v CPU in backup VM,and thus solve the double scheduling problem;3.We design and implement the COLO++ system on Linux/KVM platform,evaluations show that the packets early releasing achieves the performance of a single VM under almost certain scenarios.The active/semi-active replication model reduces double scheduling problem.