Questions regarding the drop in ExecsPerSec around March 2024

[Eulgyu Kim]

Hi all,

I was looking into the 'Graphs' section of the syzbot dashboard,
and noticed a drop in the 'ExecsPerSec' metric for the
'ci-upstream-kasan-gce' instance around March/April 2024.

(Link: https://syzkaller.appspot.com/upstream/graph/fuzzing?Instances=ci-upstream-kasan-gce&Instances=ci-upstream-kasan-gce-root&Metrics=TotalExecs&Metrics=ExecsPerSec&Months=36)

I have a couple of questions regarding this observation:

1. Is this a genuine decrease in throughput,
or is it due to a change in the measurement methodology?

2. If the drop is real, are there any specific advantages or trade-offs
that offset this performance cost?

As I am relatively new to syzkaller, I find it difficult to determine
whether this is an expected change. Any insights you could share would
be greatly appreciated.

Thank you for your time and assistance.

Best regards,
Eulgyu Kim

[Aleksandr Nogikh]

Hi Eulgyu,

On Tue, Feb 10, 2026 at 9:49 AM Eulgyu Kim <eulg...@snu.ac.kr> wrote:
>
> Hi all,
>
> I was looking into the 'Graphs' section of the syzbot dashboard,
> and noticed a drop in the 'ExecsPerSec' metric for the
> 'ci-upstream-kasan-gce' instance around March/April 2024.
>
> (Link: https://syzkaller.appspot.com/upstream/graph/fuzzing?Instances=ci-upstream-kasan-gce&Instances=ci-upstream-kasan-gce-root&Metrics=TotalExecs&Metrics=ExecsPerSec&Months=36)
>
> I have a couple of questions regarding this observation:
>
> 1. Is this a genuine decrease in throughput,
> or is it due to a change in the measurement methodology?

The measured parameter remained the same, so it's a genuine change.

>
> 2. If the drop is real, are there any specific advantages or trade-offs
> that offset this performance cost?

The change in exec/sec was a side effect of the major syzkaller
architecture change: we switched from having a fuzzing engine inside
each VM to a single fuzzing engine on the host process, which
introduced a totally different set of trade-offs across the
implementation. The new architecture seems more friendly toward larger
corpus programs, which necessarily leads to a slower execution rate.

See https://github.com/google/syzkaller/issues/1541 and all
PRs/commits that reference it.

--
Aleksandr