Work-in-progress CL:

https://chromium-review.googlesource.com/c/chromium/src/+/1407671

I haven't had time to take a lot of measurements, but initial testing looks promising.  The "ServiceWorker Preparation" time of the google sheets core stylesheet drops from 18ms to 9ms at the median.  That was only 10 samples in each mode on my fast laptop, though.  I'll do a more thorough measurement on the slow laptop Monday.

Looks good to investigate. Not sure whether SWSubresourceLoaderFactory creation is included in "ServiceWorker Preparation". It looks like "SW Preparation" is calculated by ResourceResponseInfo::service_worker_ready_time - service_worker_start_time (which are set in subresource requests in S13nSW by ServiceWorkerSubresourceLoader).
https://cs.chromium.org/chromium/src/third_party/blink/renderer/devtools/front_end/network/RequestTimingView.js?l=154&rcl=dd05131f5782853146226091e1c4d1209c47e2bb

Ah, good point.  So if this change has any effect, if any, then it will just show up in COV.

I'm unsure of what to make of the high "ServiceWorker Preparation" times, though.  It seems like the worker should be running since the navigation FetchEvent was handled.  I wonder if there is can be some delay with mojo creating the pipe for the connector to the service worker.

Maybe its because we lazily call EnsureControllerConnection() on the first DispatchFetchEvent:

https://cs.chromium.org/chromium/src/content/renderer/service_worker/controller_service_worker_connector.cc?l=33&rcl=dd05131f5782853146226091e1c4d1209c47e2bb

I'll try adding a step to my CL to eagerly call that as well.

So far I'm not really seeing any benefit here.  I'm going to try to dig into chrome://tracing to see whats going on.

#c4 doesn't matter because we get the connection in the initial subresource params.  EnsureControllerConnection() is essentially a no-op in this case.

I mean GetControllerServiceWorker() can just return the ControllerServiceWorker without calling the remote EnsureControllerServiceWorker() mojo call.

I can't measure any benefit here:

https://docs.google.com/spreadsheets/d/15HzaL0NgCZ3yynS5bzIfWilKUnOm0EZtVT2EdaYcPU0/edit?usp=sharing

I'm going to leave this open in case any thinks we should do it anyway, but it would be a bit speculative.  Feel free to WONTFIX.

To follow up here, wanderview@ found the high Preparation time was the queuing time fo the fetch event.  I think S13nSW changed the meaning of SW Preparation but we can revisit that in another bug. I'll wontfix for now, we can reopen if we discover bottlenecks in traces.