Unable to reproduce gVisor Performance Guide fio results

[Jim Junior]

Hello, I’m evaluating the storage performance of several container runtimes using fio, including runc and runsc. However, I’ve been unable to reproduce the relative performance shown in gVisor’s Performance Guide: https://gvisor.dev/docs/architecture_guide/performance/

In my results, runsc consistently performs around 5–6× better than runc for the same fio workload, which seems unexpectedly large.

I initially suspected this might be related to gVisor’s root filesystem overlay optimization. I tested runsc with:

--overlay2=none

and also with:

--overlay2=none

--directfs=false

However, disabling these features had only a relatively small effect on the results, and runsc still significantly outperforms runc.

The fio workload I’m currently using is roughly:

rw=randrw

bs=4k

size=512M

direct=1

time_based

runtime=30

I’ve attached a comparison of the results.

Is there something about gVisor’s filesystem/I/O path or fio semantics that could explain this difference? Also, is there a particular fio configuration or methodology used for the storage benchmarks shown in the Performance Guide that I should use to make the comparison more representative?

[Zach Koopmans]

Hey Jim,

Regarding reproducing the performance guide: it was generated about 6 years ago and needs a refresh. At that time, we did not have the Systrap platform (or use it as the default) which performs much better than the platform we used then.

I think it was also done using a different test harness, but the workloads (including fio) are similar.

Zach

 

--
You received this message because you are subscribed to the Google Groups "gVisor Users [Public]" group.
To unsubscribe from this group and stop receiving emails from it, send an email to gvisor-users...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/gvisor-users/4d21282c-b124-4175-afbe-98fd0c7f1731n%40googlegroups.com.

[Zach Koopmans]

Also:

"In my results, runsc consistently performs around 5–6× better than runc for the same fio workload, which seems unexpectedly large."

This should not be the case and something is wrong here. Are you using our benchmark suite?