Is there a way to help get performance stats from Alibre and figure out bottlenecks?

[sacherjj]

I'm running Alibre on a decent machine:

AMD 7950x
64 GB
3090ti

Going into a 1440p display as part of 3 display setup.

I routinely run into 2-4 second pauses when mouse wheel zooming in or selecting edit of a sketch of REALLY simple projects.

Do we have a way to enable logs or performance metrics to try to drill down on why performance pauses?

 

[Ralf]

As a first step, you can check the refresh rate of each monitor.
It should be higher than 59,9 Hz.
(A cheap Chinese HDMI cable can very quickly become a problem.)

 

[sacherjj]

Main monitor is a 144 Hz, Secondary is 75 Hz, small 1080p is 60 Hz.

We aren't talking about lagging display, but nothing changing as Alibre hangs for seconds and then catches up. I'll try to correlate to any logged issues.

 

[sacherjj]

Here is video of what occurs. I can close Alibre, re-open and continually reproduce this. Not sure if we have lazy loading of something that occurs first time I go sketching. I just seem to run into it all the time and it is annoying.

So I see about 4 seconds until anything happens, then hang when loading sketch of 14 seconds. Tons of time with a modern computer.

 

[EPowIPi]

Perhaps a profiler could help to understand what is happening in these "freezing" periods. A first very easy attempt would be the app level profiler "Very Sleepy": http://www.codersnotes.com/sleepy/
You just run it, run Alibre, click on refresh, select Alibre, click on "Profile All", switch to Alibre and move around the model as you did it in the video. And then look at the result to see whether there is time spent in unusual functions.

Here how this looks for me, without any freezes:


If that doesn't help then you need a system level kernel profiler that also records in more detail what is happening in system libraries, drivers (e.g. GPU driver) and in the kernel and all processes at once. One would be the Windows Performance Analyzer (https://learn.microsoft.com/en-us/windows-hardware/test/wpt/windows-performance-analyzer). This is a little bit more involved. There are various tools for working with that (=recording the traces to watch in WPA) in Visual Studio, but also open source tools like this: https://github.com/google/UIforETW. This site also contains links how to use this.

 

[sacherjj]

I've done mostly Linux and Rust development so more familiar with Linux sys calls. Been over a decade removed from Windows development.

 

[sacherjj]

Also captured the same action that doesn't hang the second time to compare.

 

[EPowIPi]

Interesting that our two traces are so different. The main difference I can see between your two traces is that in the first case WaitForMultipleObjects has a much higher waiting percentage time. And exclusive is almost the same percentage as inclusive (I think that means that it waits for resources that are not occupied by any Alibre thread but by something external).
But I fear you will have to open a support ticket, I guess we cannot solve this from this information alone.

 

[sacherjj]

Tried a Windows Performance Reader and crashed when loading for analysis.

 

[EPowIPi]

There is a thread where somebody seems to have the same problem: https://www.reddit.com/r/WindowsHelp/comments/1hp5jd2/_/m4ifmis Apparently there are different versions of WPA and traces. They propose to use the WPA version from the app store. Or set verbosity to light as the differences seem to be in level detail(?)

 

[Ex Machina]

Also, make sure that Alibre is set up to use the GPU and not the APU... Just thought I'd say that.

 

[Ralf]

The file should also be saved locally, i.e. not with Onedrive, or other clouds, network storage...

 

[sacherjj]

The file should also be saved locally, i.e. not with Onedrive, or other clouds, network storage...

It was captured in RAM and then saved locally.

 

[sacherjj]

Also, make sure that Alibre is set up to use the GPU and not the APU... Just thought I'd say that.

Their process of changing the path of Alibre with every upgrade makes you go in and setup Performance mode each upgrade. I just did that again with 28.2.

This did help, as the 4 / 14 pause turned into 2 / 7. But still much more than I would expect, unless lazy loading or processing something large at the first load of Sketch components.

 

[Ex Machina]

The file should also be saved locally, i.e. not with Onedrive, or other clouds, network storage...

To that point. I save everything on OneDrive which gives me versioning after every save even before PDM. Never had an issue...

 

[DavidJ]

From 28.0.2.28126, the setting of graphics adapter SHOULD work much better and not need intervention each time. Maybe a little soon to be entirely confident of that...

 

[DavidJ]

Cloud storage is another potential variable. It seems to work well for some customers, but we've also seen tales of woe. Probably very dependent on your internet connection quality.

 

[Ralf]

It was captured in RAM and then saved locally.

Try to save locally and open again from there. Is there a difference?

 

[sacherjj]

I'll try some of the options here and report back. Have to do real work now.

 

[DavidJ]

Anything that is repeatable in terms of triggering delays should be reported to support. Obviously if you get any clues from your own analysis, provide that information too...

Development can only investigate (and hopefully fix) when they have cases/examples to study.