Profiling info for the network evaluator similar to the profiling info for processors

[r-englund]

I have wanted to have a nice way in Inviwo to see how long the network takes to evaluate, similar to what we have for each processor:

I have quickly (and ugly, codewise) hacked a prototype, which can be seen below. It is a widget added to NetworkView, uses the ProcessorNetworkEvaluationObserver to detected when evaluations start and stop.

Before putting in more work into it I want some feedback from you.
Right now it uses the StopWatch icon (same as the tool button to reset the processor profiling). ATM it is way too small and has a very bad contrast towards the background. Things that can be done here are, for example: Make it bigger and change color for more contrast; Use text with the current number of evaluations (like the count we have on the processor graphics items); Both text and icon, maybe bigger size.

[r-englund]

Zoom in on current icon without any modifications and no tool tip

[r-englund]

@martinfalk @ResearchDaniel @petersteneteg

[ResearchDaniel]

I guess this is too much work but how about a tree-view with all the profiling measurements? Would be reached through the menu instead...

[r-englund]

I guess this is too much work but how about a tree-view with all the profiling measurements? Would be reached through the menu instead...

I have also thought about that, to have a sortable table view of the profiling measurements for each processor, also including (optional) gpu times. But that requires way more work and solves a different issue

[ResearchDaniel]

Ok, then a "drop-down" instead of pop-up would make sense:
https://material-ui.com/components/expansion-panels/

I would place it below the search bar in the top right and make the background the same as our properties.

[petersteneteg]

I have a wip branch that does much better profiling then what we get from the cpu callbacks... it also plots stats over time

[r-englund]

Those features you guys suggest are good, but I do not think it makes this feature obsolete.

Me and @martinfalk talked about it and the idea is to move the "icon" down to the status bar and show the tooltip on hover. Something like this (with proper font, not my sloppy handwriting)

[r-englund]

So the question is, are there any interest in my crating a PR for this? Or should I keep it in my own module?

[martinfalk]

I am pretty much in favor of the status bar solutions for some quick feedback. Detailed profiling by @petersteneteg would complement this and is welcome as well :-)

[r-englund]

I am pretty much in favor of the status bar solutions for some quick feedback. Detailed profiling by @petersteneteg would complement this and is welcome as well :-)

The only odd thing with the status bar solution is that the status bar is part of the console widget and not the main window.

[petersteneteg]

@r-englund I noticed that when I added the treadpool info thing. Should probably be moved into it's own thing some how

[tirpidz]

It would be great to have a visual feedback in the graph itself: processor widgets changing colour when they take more time than usual or when the take a big chunk of the whole network evaluation time.

[tirpidz]

An other way, would be to use the chromium devtools: https://developers.google.com/web/updates/2016/12/devtools-javascript-cpu-profile-migration From the look of it, there is already a lot of tooling to profile & you can record/load a profile with a .json format.

[{"args":{"name":"CrBrowserMain"},"cat":"__metadata","name":"thread_name","ph":"M","pid":23703,"tid":23703,"ts":0},
{"args":{"name":"CrRendererMain"},"cat":"__metadata","name":"thread_name","ph":"M","pid":28787,"tid":1,"ts":0},
{"args":{"name":"Chrome_IOThread"},"cat":"__metadata","name":"thread_name","ph":"M","pid":23703,"tid":23722,"ts":0},
{"args":{"name":"VizCompositorThread"},"cat":"__metadata","name":"thread_name","ph":"M","pid":23738,"tid":23940,"ts":0},
{"args":{"name":"CrGpuMain"},"cat":"__metadata","name":"thread_name","ph":"M","pid":23738,"tid":23738,"ts":0},
{"args":{"name":"SGI_video_sync"},"cat":"__metadata","name":"thread_name","ph":"M","pid":23738,"tid":23942,"ts":0},
{"args":{"name":"Chrome_ChildIOThread"},"cat":"__metadata","name":"thread_name","ph":"M","pid":28787,"tid":4,"ts":0},
{"args":{"name":"Compositor"},"cat":"__metadata","name":"thread_name","ph":"M","pid":28787,"tid":7,"ts":0},
{"args":{"name":"ThreadPoolServiceThread"},"cat":"__metadata","name":"thread_name","ph":"M","pid":23703,"tid":23720,"ts":0},
{"args":{"name":"Chrome_DevToolsADBThread"},"cat":"__metadata","name":"thread_name","ph":"M","pid":23703,"tid":25997,"ts":0},

A quick way to leverage this tool would be to have each processor the events you want to profile & dump it into a json file with the proper format & use the devtools in chrome. My guess is there is way more options/granularity already available waiting to be used :-)

[ResearchDaniel]

@tirpidz is the profiler based on a web component? In that case it would even be possible to feed the info through a data frame directly into a webpage using the web browser. Not as integrated but another option perhaps

[tirpidz]

not 100% sure on how it works in chrome, I just hit: ctrl+shift+i inside chrome to bring it on. My guess would be it is web based & open source ;-) A front-end guy would have a much better knowledge of it.