Prof. Dr. Frank Bellosa, Marc Rittinghaus

Abstract:

Checkpointing a virtual machine (VM) in high-frequency intervals poses a challenge for checkpointing mechanisms. The stop-and-copy approach suffers from high downtime during checkpointing. That impedes its use in interactive scenarios, for example in workloads that utilize the network or interact with a human user. The pre-copy approach helps to reduce downtime by copying the VM’s memory during its execution in multiple copy rounds. However, it is not able to reduce downtime reliably when the VM’s writeable working set is too large.

This thesis aims to solve this problem by copying the VM’s memory concurrently to its execution. Our approach utilizes copy-on-write to preserve the consistency of the virtual machine memory checkpoint image. In our evaluation, we have proven this mechanism to provide a predictable, almost constant downtime of 6ms during a kernel build with a two-second check- pointing interval. We have determined that our implementation yields a small constant overhead while copying pages and is therefore able to keep up with a checkpointing interval length down to 16ms.

BibTex:

@mastersthesis{boehr15copyonwrite,
 author = {Nico Boehr},
 title = {Evaluating Copy-On-Write for High Frequency Checkpoints},
 type = {Bachelor Thesis},
 year = 2015,
 month = sep # "30",
 school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
}