app using "excessive" CPU

[mzimmers]

Hi all -

I have a small app whose display looks like this:

The model grows to about 50 columns (most of which are hidden) and 20 rows -- certainly not huge. And in the first few seconds, the model rarely gets updated.

Multiple users have complained about it using too much CPU (about 11% on my i3 when seemingly idle).

Given that this is running on Windows, I can't use the profiler built into Creator. Any tips for trying to determine where all this CPU is going?

Thanks...

[fcarney]

I would assume that for some reason it is receiving a lot of events or a thread is busy doing "something". Possibly a combination. Is there a way to tap into and print out the events the app is receiving while idle?

[mzimmers]

Good assumption - I'm getting thousands of events per second, all like this:

15466 "15:46:02" QEvent::Paint
15467 "15:46:02" QEvent::UpdateRequest
15468 "15:46:02" QEvent::Paint
15469 "15:46:02" QEvent::UpdateRequest
15470 "15:46:02" QEvent::Paint
15471 "15:46:02" QEvent::UpdateRequest
15472 "15:46:02" QEvent::Paint
15473 "15:46:02" QEvent::UpdateRequest
15474 "15:46:02" QEvent::Paint
15475 "15:46:02" QEvent::UpdateRequest

But what could be causing this? My model isn't changing (certainly not that quickly) and there's no user input, or resizing going on. The app, from the user's perspective, is truly idle.

[fcarney]

This is guessing time. I am good, but not that good. But I will hazard a guess. Something is "wiggling" and it most likely caused by something that controls layout. Something adjusts, something else adjusts it back.

I do most of my work in QML and when this happens it is usually detected as a binding loop. I am unsure if there is anything like that in QWidgets.

[gde23]

You can start the debugger and after start randomly click on pause.
Since the program will be inside Paint or UpdateRequest with a high probability you will get there quickly and then you can follow the call-stack to see what is causing it.

[mzimmers]

@gde23 that's a clever idea, and I tried it. Unfortunately, whenever I pause the debugger, the call stack only displays routines pertinent to handling the pause:
Great idea, though.

[SGaist]

Hi,

Are the columns automatically resized to content ?

[mzimmers]

@SGaist not explicitly. I do this:

    qhv->setSectionResizeMode(QHeaderView::Stretch);

But disabling that line doesn't reduce CPU noticably.

[JonB]

@mzimmers
Sit down, grab a coffee & a pack of cigs. Start binary-chopping-commenting-out half your code at a time. It can't be that hard to get to the point where these update requests do not happen... :) For example, first get rid of all hidden columns. Then reduce to fewer columns than the width of the window. With 2 rows. Then no rows and no columns --- does it still happen then?!

What are you using which gives you those events printed out?

[mzimmers]

@JonB OK, I'll belly up to removing the hidden rows (I was hoping to avoid that).

I've already sized my widget so that everything is fully displayed.

I even disconnected the signal that tells the model to update. My tableView is completely empty. And I'm still getting the CPU usage.

Here's what I use to show me the events:

bool KeyPress::eventFilter(QObject *obj, QEvent *ev)
{
    bool rc = false;
    QEvent::Type type;
    static int count = 0;

    type = ev->type();
    qDebug() << count++ << QTime::currentTime().toString() << type;

    rc = QObject::eventFilter(obj, ev);
    return rc;
}

It may be noteworthy that disabling this filter entirely doesn't help.

[mrjj]

Hi
Could you not add a paintEvent handler where you simply call base class but
then add a break point and that way maybe see how its triggered ?

[JonB]

@mzimmers said in app using "excessive" CPU:

My tableView is completely empty. And I'm still getting the CPU usage.

Which is the sort of thing which is interesting, so you don't want to waste time looking at resizes. Nonetheless, the events indicate widget updating?

I had forgotten about your KeyPress and its eventFilter. Get rid of it completely from your code! You don't want to be messing about with your key press stuff from your other thread. Now what's the speed? It ought be still slow, but you never know.

Then get rid of all your own "event filters". Reintroduce only one at the very top level to debug monitor whatever you need to.

[mzimmers]

@mrjj I can do something like this:

void Widget::paintEvent(QPaintEvent *event)
{
    QWidget::paintEvent(event);
}

But I don't know what kind of meaningful information I can extract from the event argument. (According to the docs, paintEvent() isn't a slot.)

[mzimmers]

@JonB I've gotten rid of all filters, all hidden columns, and have disconnect model updates after I create two rows. I also disabled stretch mode, and expanding policy. Still using 11% CPU.

This might be a good time to point out that I "stole" the code for this app from another project, which doesn't have this problem. For the life of me, I can't see what I added that causes this.

[JonB]

@mzimmers
Try not running your application, or switch the machine off ;-)

Given that this is running on Windows, I can't use the profiler built into Creator

Does your MSVC/MinGW compiler come with a usable profiler, nothing to do with Qt?

In your paintEvent() override @mrjj had in mind for you to try a debugger breakpoint there and look at the stack trace. You will doubtless need to perhaps put a delay/count on the breakpoint, or set it while the app is in the middle of running. I don't actually know whether paint events will show anything interesting on the stack.

You can use the top-level eventFilter() to examine most of what's going on, if you have to :( Debug every event's type to file/debug output window for a couple of seconds into your "idle". Search the output, you're probably interested in what you see just before the updating starts, or as it goes along.

[mzimmers]

@JonB using MinGW, and I don't think it has a profiler. I could look at 3rd party products, but it's probably easier just to have a co-worker build this on Linux, and then run the profiler. (Hopefully the problem will occur on Linux!)

Here's the stack trace you mentioned:

Maybe you can see something useful in it; I can't.

[JonB]

@mzimmers
An expert may have something to say about the traceback. All I know is it would be nice to know which the QWidget is. Wait, this code is yours for Widget? So which of your widgets is it? If you break more than once, is it always the same widget?

I can see QCoreApplication::sendSpontaneousEvent(). You sure you're not "wiggling"? :)

(Hopefully the problem will occur on Linux!)

Nope... ;-)

[mzimmers]

@JonB yes Widget is my oh-so-creative name for my main QWidget class, which is the only QWidget the app uses (unless you push one of the buttons). So yeah, it's always the same Widget.

Someone else mentioned wiggling, but I'm not sure I know what it means. The problem occurs even when the app loses focus, though.

And, if you're confident this won't happen on Linux, then maybe the problem isn't in my code space...

[JonB]

@mzimmers said in app using "excessive" CPU:

And, if you're confident this won't happen on Linux, then maybe the problem isn't in my code space...

Noooo, I put a wink --> ;-) <--
I absolutely do not know whether it will repro under Linux, if you're lucky it will, I just meant sod's law it won't!

[mrjj]

@mzimmers
Hi
That stack trace is for

1. starting app
   2: add break point
   3: let it loose focus to see the issue

and not just set at startup so we ssee the first paint when it becomes visible?
Just asking to be sure. Not seeing anything special besides maybe the sendSpontaneousEvent