Hi, On Tue, Jul 18, 2017 at 06:55:12PM +0200, Christophe de Dinechin wrote: > Interestingly, I just discovered something I did not really expect: a > hot CPU is more than enough to make the spice client lack “oxygen”, > even if nothing else is running. > > If I start under light load and keep running that way, it can run a > long time without a problem. The queue length typically runs up and > down a little, but it does not diverge. See > http://blackbox.dinechin.org/images/170718-NotDiverging.png. > > There is even plenty of CPU left to converge back if we temporarily > accumulate things in the queue: > http://blackbox.dinechin.org/images/170718-Reconverge.png. > > However, if the machine gets hot, then it does not reconverge, because > the machine spends over 50% “in the kernel” just trying to cool down. > Starting the client under these conditions, it rapidly diverges: > http://blackbox.dinechin.org/images/170718-WithHotCPU.png. Isn't it expected as the CPU lost 50% of its processing time in order to cool down? We can think about different ways to workaround this kind of issue. Dropping the queue as you did in a different patch is a good approach for peaks like this, I think. If those situations happens too often it would be a good signal that current stream is too much for the client to handle and we could lower quality/fps, increase I-FRAME, etc. in order to make it less demanding for the client. Still, it should be much better if you were doing hw decoding, I guess. > Also, without enough I-frames to recover, when this starts happening, > we are a bit stuck. This is all documented in more details here: > http://blackbox.dinechin.org/170718.html. Cool charts > Tons of messages that look like: > > channel-display-gst.c:266: [783802 1768.364819] spice_warning: got an > unexpected decoded buffer! First time seeing it, cool! > So I think we will need the mechanism we discussed with Uri, where we > send something back to tell the server we need an I-frame. Or we > close/reopen the channel. Definitely! We should start discussing about it... I think it will require a new/improved protocol. Cheers,
Attachment:
signature.asc
Description: PGP signature
_______________________________________________ Spice-devel mailing list Spice-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/spice-devel