Enhancing System Checkpoints for Seamless Fault-Tolerant Live Virtual Machine Migration

Abstract

Abstract Fault-Tolerant Live Virtual Machine migration is utilised in Cloud Computing (CC) to ensure smooth and risk-free migrations, ensuring the data and operations of the running VMs are kept intact. To mitigate the risks of the failures that could arise during the migration periods, that could lead to loosing the VM, due to the VM’s data and states being split among the source/destination servers and the network cables during transit, checkpoints/snapshots are utilized. Incremental checkpointing utilizes an iterative fashion where the memory content is checkpointed in several iterations to avoid multiple bulky VM image-level checkpoints being captured and stored. It captures the main memory content in a series of iterations, where the first checkpointing iteration captures all the memory pages available at that instance, and the successive checkpointing rounds capture only the modified pages when compared with the earlier iterations. This research optimises the incremental checkpointing approach by making the memory capturing more fine-granular. The Mem-Finer model developed, uses a granular methodology to successfully extract the actual modified content within a modified page without repeatedly sending the whole 4096-byte page redundantly. Additionally, three lossless compression algorithm-enabled models, Gzip, Lz4, and Zstd, target compressing the modified pages efficiently before sending them to the checkpoint stores. The four models developed showed mixed results where the Mem-Finer model and Lz4 showed drastic average page content reductions of 51-69% and 69-72% respectively, while showing similar and even lesser overhead patterns in migration times, downtimes, checkpoint replication times, performance impact and recovery times, depicting that it successfully enhances the baseline incremental checkpointing approach by reducing the network bandwidth occupied by each VM, ultimately increasing the productivity and cost efficiency of cloud environments. Eventhough Gzip and Zstd models showed good page content reductions of 78-79% and 50-52% respectively, it was observed that Gzip showed poor performance in memory write-intensive environments, and Zstd showed poor performance in all the environments tested. The results highlight the overall better efficiency when using the Mem-Finer and Lz4 models, and the poor efficiency showed by Gzip and Zstd under the specific workloads.