Prof. Dr. Frank Bellosa
Dr. Marc Rittinghaus

Full system simulation is a flexible tool for dynamic analysis of computer systems. However, it causes a large slowdown for most workloads. This slowdown can also change the behavior of the observed system. SimuBoost addresses this by recording the workload in a hardware-assisted virtual machine and replaying it in a full system simulator. Therefore, the behavior of the workload is not affected by the large slowdown of a full system simulator. To speed up the analysis, continuous checkpointing is used to create independently analyzable segments. These segments can be analyzed by multiple workers in parallel.

However, the continuous checkpointing introduces a new overhead source into the recording process. An additional replay stage can be used to separate the recording from the checkpointing. So far, SimuBoost has only implemented a replay engine for full system simulation. However, running the checkpointing in a full system simulation with its associated slowdown would cause significant delays in the parallelized analysis. Therefore, a replay for hardware-assisted virtual machines is necessary.

We implemented the necessary replay engine and compared its run time with the run time of the recording. Ideally, the difference should be small so the move to the stage does not affect the checkpointing. We measured that in most workloads the run-time overhead is less than 5%.

BibTex:

@bachelorthesis{schlumpp19systemsimulation,
  author = {Marco Schlumpp},
  title = {Towards Three-Stage Parallelization of System Simulation},
  type = {Bachelor Thesis},
  year = 2019,
  month = oct # "16",
  school = {Operating Systems Group, Karlsruhe Institute of Technology (KIT), Germany}
  }