kMemvisor: flexible system wide memory mirroring in virtual environments

  • Authors:

  • Bin Wang;Zhengwei Qi;Haibing Guan;Haoliang Dong;Wei Sun;Yaozu Dong

  • Affiliations:

  • School of Software, Shanghai Jiao Tong University, Shanghai, China;School of Software, Shanghai Jiao Tong University, Shanghai, China;School of Software, Shanghai Jiao Tong University, Shanghai, China;School of Software, Shanghai Jiao Tong University, Shanghai, China;School of Software, Shanghai Jiao Tong University, Shanghai, China;Intel China Software Center, Shanghai, China

  • Venue:
  • Proceedings of the 22nd international symposium on High-performance parallel and distributed computing
  • Year:
  • 2013

Quantified Score

Hi-index 0.00

Visualization

Abstract

Today's commercial cloud service providers require the availability with an annual uptime percentage at least 99.95\%. While memory errors become norms instead of exceptions with the increasing memory's density and capacity in cloud applications. Thus, uncorrected errors from DRAM can be a significant source of system downtime. To address this increasingly important concern, both hardware and software memory mirroring technologies are studied nowadays to provide memory high availability. However, hardware solutions like mirror memory, which uses doubled chip, need dedicated and costly peripheral hardware. While existing software approaches, i.e., virtual machine's checkpoint technology, reduce the expense but incur the high overhead in practical usage. In this paper, we present a novel system called \emph{k}Memvisor to provide system-wide high availability memory mirroring. It is a software approach achieving flexible multi-granularity memory mirroring via virtualization and binary translation technology. Specifically, kMemvisor first creates backup space of the same size of the specified memory for applications or virtual machines. We can flexibly set memory areas to be mirrored or not mirrored from application level to system-wide. Then, all memory write instructions in the native memory space are captured and instrumented by mirror memory write instructions to synchronize the data in backup space. Furthermore, this instruction level memory synchronization reduces backup overhead and lowers the probability of data loss compared with traditional software approaches. So kMemvisor could use data from the backup space to recover when memory failures happen. The results show that kMemvisor causes 55% overhead in the worst case of system-wide high availability and 30% average for the real world applications, which outperforms the state-of-the-art software approaches even in the worst case.