memscroller stresses the html5 client

[totaam]

Split from Xpra-org/xpra#2231, see Xpra-org/xpra#2231#comment:9.

Damage latency goes up to 7 seconds with chrome!
Average is also too high.

[totaam]

2019-08-11 07:21:07: antoine uploaded file memscroller.png (22.2 KiB)

memscroller latency

[totaam]

More test data showing the memscroller outlier problem in Xpra-org/xpra#2231#comment:12.

[totaam]

As per Xpra-org/xpra#619#comment:29, the problem occurs only with NODELAY=0.
My guess is that this buffers too many packets in the kernel's socket output buffer, allowing things to spiral out of control.

[totaam]

2019-08-19 17:32:10: antoine uploaded file memscroller-server-rss.png (59.2 KiB)

memscroller also uses too much memory

[totaam]

As per comment:3 and Xpra-org/xpra#2231#comment:12, the default configuration handles this well enough that this can wait for the next release.

[totaam]

It's a lot better than it used to be, but I'm still seeing problems with the nvjpeg encoder: we're doing a stable 30fps for a while then backing off, then back to 30fps...
This may help:
Xpra-org/xpra@b0f2859

But ideally, the html5 client would be able to sustain the 30fps framerate without this jitter.
Why doesn't it?