Adding Blob Perf Tests

[Daniel Murphy]

Hello!

I'm trying to add some telemetry tests to measure the performance of Blob storage.  Since it's not something that's used very often yet in pages, I'm creating a custom page set and static page to exercise the various parts of blob storage.  I'm also adding tracing to the parts of the system I want to measure.

I'm trying to figure out what I need to do to hook this all up in a perf test.  I imagine this will be similar to IndexedDB, but I'm not sure if I need to add my own measurement to record the in-page timing results.

Would you be able to outline what else I need to do?

Thanks!

Daniel

[Daniel Murphy]

Gentle ping, +sullivan

[Nat Duca]

Ideally you'd just use timeline_based_measurement and then insure that one of the metrics in timeline_based_measurement reflects what you're trying to measure. The thread times metrics seem promising: we track total wall time and total cpu time --- if blob storage gets faster or slower, then these numbers should go up and down, right?

Try following the tough_x_sites convention in the pagesets dir for your new pageset. Making synthetic pages sounds good.

[Daniel Murphy]

I'm a little scared that it won't give us good data, but I guess it could work for now.  In that case, should I just have the pages automatically run 'benchmarky' things with our system, and then have one page per benchmark?

[Nat Duca]

Hehe, fair enough! What's your ideal data that you'd like to get from a measurement? You might be able to extract specific traces as well --- a la layout.py metric --- the key thing is using timeline based measurement.

[Daniel Murphy]

Quick overview of blobs:

* creating a blob is synchronous for sending memory to browser 

  * all memory-backed, no disk, and we have a global cap

* reading is a network request, all from memory on IO thread.

We are changing to:

* creating a blob is asynchronous and can be disk backed based on the size and the current storage size

* reading can block on pending data from the renderer.

Ideally we would want specific timing information on

* How long it takes to register a blob

* How long individual items take to add

* How long it takes to read a blob

* How long for the whole round trip, construct -> read

I added tracing for most of these, except the round trip.  I imagine if we make specific pages to benchmark each of these, we can use the measurements that you mentioned instead?