[Spice-devel] Processing of rendering operations in separate thread

[Yuri Benditovich]

Hello Javier,

After implementing the pushing thread in current qxl-wddm-dod I measure the
CPU consumption in PresentDisplayOnly call and in the thread on pushing
drawables to the device. The results show that the time to push drawables
is negligible in relation to time of copying dirty rects to the device
memory (in average the proportion is ~ 1/500) and in typical case the
pushing thread serves only single 'present' operation and then waits for
next operation.
I tried in on Win10 with 2-3 G memory and 1-2 CPUs with regular user
activity (opening windows, redrawing, scrolling etc)

Do I miss something?

Thanks,
Yuri


On Fri, Nov 25, 2016 at 11:11 AM, Javier Celaya <javier.celaya at flexvdi.com>
wrote:

> Hello Yuri
>
> El vie, 25-11-2016 a las 01:08 +0200, Yuri Benditovich escribió:
>
> I'm porting to [qxl-wddm-dod] set of flexvdi changes
> related to execution of 'present display only' events
> in separate thread. There are 2* questions below* I'd like to ask and know
> your opinion.
>
> I see there 2 aspects:
> - reliability
> - performance
>
> *Reliability:*
> I see in flexvdi mailing list existing report of
> BSOD upon system shutdown. Possible cause is lack of
> synchronization between system flows, hardware availability and worker
> thread state (last patch in flexvdi 'Terminate working thread on exit' introduces
> termination procedure but nobody calls it, as I can see)
> The lack of synchronization may cause also races in
> power management flows and (possible) on changing
> operating mode.
>
> *Question 1:*
> *Do you have some additional recommendation which*
> *flows shall be specially checked for races with*
> *rendering thread?*
>
>
> Unfortunately, the truth is, we have not thoroughly tested our code to
> remove these races yet. The clients this driver was intended for are still
> stuck using Windows XP/7, and our development is stalled. So, I cannot
> think of any situation you should check that you do not know about yet.
>
>
> *Performance:*
> It looks like the change should not affect total CPU consumption for
> the rendering, it splits more or less the same operations over
> 2 different threads. It is still possible that the change can improve
> common user experience due to faster indication of operation completion
> to the OS.
>
>
> We were not trying to reduce total CPU consumption. After all, the driver
> just copies rects from main memory to VRAM and passes them to the spice
> server; there is little to reduce there. Rather, we tried to increase the
> throughput of graphic operations, by not locking the DirectX subsystem
> while we wait for the spice server to accept new drawables. That is, we do
> not mind using more CPU if that results in painting faster.
> On the other hand, I was thinking that maybe we could get the DirectX
> subsystem to provide the rects already in VRAM if we described it as a
> linear memory segment on driver initialization. In that way, the copying
> operation could also be removed. However, I am not sure if this actually
> works or even how to do it, it is just an idea.
>
>
> *Question 2:*
> *Do you have some ideas how to make quantitive*
> *evaluation of this possible **improvement of user experience? *
>
> I think about:
> - finding scenarios when we receive rendering calls (PresentDisplayOnly)
> when the worker thread is still processing previous operation. If they
> exist this can mean that some bottleneck solved in GDI.
> - writing or getting tool that loads the graphics
> adapter by heavy operations (like continuos moving of window / scrolling
> etc) with CPU consumption measurement
>
>
> We used a simple tool to measure the performance: it creates a window and
> continuously issues WM_PAINT events where the full background is filled
> with color, then measures the number of processed events per second (not
> CPU). It is quite naive, but it provides a good starting reference, since
> the tool, with the XDDM QXL driver in Windows 7, outputs almost twice as
> much paint events as executing it in Windows 8 with the WDDM QXL driver.
> There are other measurements you can try to obtain, like how much time does
> it take until a paint event gets to the spice server queue, ready to be
> sent to the client (although I'm not sure how to measure it). This delay
> affects the user perception of performance.
>
>
>
> Please share your thoughts.
>
> Thanks,
> Yuri
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/spice-devel/attachments/20161130/a20116e6/attachment-0001.html>