Virtual machine online migration method based on hardware acceleration

Abstract

The invention discloses a virtual machine online migration method based on hardware acceleration. The method comprises the following steps: firstly, the performance influence of different loads on different migration models is analyzed; two basic strategies for optimizing the migration performance are provided by utilizing the obtained conclusion; in other words, different migration model strategies and load-aware pre-copy migration strategies are used for different loads, and a set of complete and efficient virtual machine migration method capable of adaptively selecting favorable migration models according to the load type of a work set is provided. Secondly, acceleration is performed by using a hardware virtualization extension function in load analysis, so that more efficient memory write-in access sequence tracking is realized; thirdly, a CLOCK + algorithm with better overhead and performance is proposed to predict a memory write-in work set; according to the method, the working set information of the upper-layer load is accurately sensed and identified, and the favorable migration model is adaptively selected according to the type of the working set load, so that the performance influence in migration is reduced as much as possible.

Description

technical field [0001] The invention belongs to the field of data center management, and in particular relates to an online virtual machine migration method based on hardware acceleration. Background technique [0002] Virtual machine migration technology mainly refers to the transfer of a virtual machine and its running status from one computing node to another, ensuring that the virtual machine can continue to provide services at the destination when the source computing node is shut down due to failure or maintenance. Virtual machine migration technologies mainly include downtime migration and online migration. Because virtual machine downtime migration will cause long-term service interruption, the industry and academia mainly use the virtual machine online migration mechanism. [0003] In terms of virtual machine online migration, a method based on pre-copy is currently the most mainstream solution to realize online migration. The method copies and transmits the runnin...

AI Technical Summary

Problems solved by technology

[0011] 1. For all the above methods, hardware acceleration is not used when tracking memory history access during the migration process (only tracking memory access can ensure that the data modified at the source end can be retransmitted to the destination end during the migration process), and hardware acceleration will be introduced without hardware acceleration. A large number of virtual machine exit events (VM-exit), which will bring a lot of overhead to the guest virtual machine

[0012] 2. Document 1 and Document 2 only use the pre-copy method. On the one hand, it will generate more invalid transmission and bandwidth consumption. On the other hand, it will cause longer migration time and affect migration reliability.

[0013] 3. In the pre-copy method, literature 1 and literature 2 use queues to maintain system access to hot data (working set data). On the one hand, the runtime maintenance cost of this queue is very high, which will bring a lot of overhead to the client virtual machine ; On the other hand, the size of the queue is fixed, and the size of the working set cannot be flexible, which limits the optimization effect

Although this method reduces the retransmission of dirty data, it only simply mixes pre-copy technology and post-copy technology, and does not consider the performance advantages and disadvantages of different loads for different copy methods. Therefore, a large number of requests are made from the source host in the early stage after migration data, which brings high latency and severe performance degradation to virtual machine applications

[0015] 5. Document 4 uses the Hidden Markov Model (HMM) to predict the probability of each memory page becoming dirty in real time, but training HMM in real time will cause a lot of additional overhead

Images