Skip to content

Latest commit

 

History

History

fig-11_flexos-alloc-latency

README.md

FlexOS Latency Microbenchmarks

Estimated prep. time Estimated runtime
0h 35m 0h 04m

Overview

Gate and allocation latency microbenchmarks. FlexOS is compared with Linux (with and without KPTI).

Measurement workflow

This experiment is somewhat special, as it requires one measurement with KPTI disabled on the host system. This requires changes to the kernel command line and a reboot. The workflow is therefore slightly different, but we tried to make it as streamlined as possible:

  1. Prepare the experiment as usual with make prepare
  2. Run the first numbers as usual with make run
  3. Disable KPTI using ./toggle-kpti.sh off
  4. After the reboot, run make run again
  5. Re-enable KPTI with ./toggle-kpti.sh on

As soon as you are done with step (4), we recommend that you immediately proceed to step 5 to reset the machine to its initial state.

Note (especially to ASPLOS'22 AE reviewers): disabling KPTI will affect all other measurements, not only Linux for this figure. Make sure to re-enable KPTI as soon as you are done with this measurement. Cooperation among reviewers for this benchmark is recommended.

Plot

The latency plot contains some hardcoded data. If you want to generate a plot for your new run, you will need to edit the plot script. The values that you will want to adapt are the labels (and potentially their position).

Troubleshooting

  • Problem: Every once in a while, EPT results increase to ~700 cycles (instead of ~500 expected).

    Solution: It is not entirely clear why this happens. We suspect a combination of unlucky scheduling of vCPU threads (not pinned in this specific case due to a missing implementation in qemu-guest), and cache effects. In our measurements this was a rather rare symptom and we therefore considered it as minor. We recommend to simply restart the benchmark in this case.