Enable PerfView To Export SpeedScope Data Not Per Thread But All Threads Aggregated

[Alois-xx]

I have whipped together a small change to export SpeedScope Data not thread wise but as an aggregate.
That makes it much easier to find CPU bottlenecks which happen in the same methods split accross many threads.
Currently the Speedscope output is not that useful if you have many threads in your application.

• You loose then the timeline view which is not that useful anyway if you can check only one thread at a time.
• Left Heavy and Sandwich work as expected

For the export menu we would then need an additional option or should that become the new default?

• Speed Scope (*.speedscope.json)
• Speed Scope All Threads (*.speedscope.json)

Any opinions?

[brianrob]

@adamsitnik

[adamsitnik]

Hi @Alois-xx

As you have pointed out, speedscope does not support showing data from multiple threads using a single view mostly due to the Time Order view limitations (jlfwong/speedscope#117). But, if you are not interested in Time Order, why would you use Speedscope while you can see the aggregated data in PerfView?

That makes it much easier to find CPU bottlenecks which happen in the same methods split accross many threads

You should be able to see these methods in the "By Name" view in PerfView (the data is always aggregated).

If you want to have "aggregated" Flamegraph you can use the [fold threads] Thread -> AllThreads grouping (please see #980 for more)

If you want to have something like "Time Order" view, but with all threads, you can use the "Chromium" trace file exporter (#1113).

Please let me know what you think about the existing options.

[Alois-xx]

Actually I am not using PerfView but a command line tool which automates the loading of an ETL, selecting a process, loading symbols or using cached ones. By adding a Thread xxxx at the bottom of each stack I get a view which shows e.g. how much threads have been spinning or if only one thread did all the work which is pretty useful in heavily multithreaded applications.

What I am most interested is the Left Heavy graph:

By doing this I also get a nice per Thread Cost View in Sandwich

When I click on that Thread there I get all methods called in this thread and their costs. That is a much more useful
view on a multithreaded application.

I understand if you want to keep the view as is. It is just a suggestion. I am going to keep my code anyway because PerfView has not yet long path support which makes it unusable to reliably load symbols from deep path names and I find a command line extraction tool easier to work with my existing tool chain.

With regards why not using PerfView:
I think Speedscope is a great vehicle to show call stacks to non ETW guys by sending the zipped .speedscope.json file via mail with a link to the speedscope app and they can view the data right away without any additional tool installation.
Usually it is only about 4 MB which passes any attachement filter.