I really don't understand at all what you're proposing. How is the quality adjusted in response to BWmax? It seems to converge way too slowly, and I don't like the fact that it will send a higher-quality frame every 30 seconds (that will cause a stutter.)
Being the one that will have to maintain this, if the solution is inelegant, it's not something I'm going to want in the code, because it will just cause problems for me down the road. I really am starting to think that an elegant solution for this is not possible given the constraints of the VNC environment. It would require some fundamental modifications to the protocol. On 5/11/11 10:39 AM, Arthur Huillet wrote: > On Tue, 10 May 2011 14:09:41 -0500 > DRC <dcomman...@users.sourceforge.net> wrote: > >> An accurate measurement of BWmax is critical >> here, though, and I don't really trust that there is a way to measure >> that accurately. > > After trying TigerVNC's BWmax measurement, and trying to do more or less the > same in TurboVNC, I am not too confident with the timing method to know BWmax. > > I'm just throwing out an idea: perhaps we can determine BWmax empirically. > > 1-Determining BWmax automatically > > My idea is to start TurboVNC with a high quality. When the network is the > bottleneck, it will be loaded at 100% in this setup, so every two seconds > apply > if (bw > max_bw) > max_bw = bw; > > (using some smoothing is a possibility but I try the simplest first) > > This gives a rough estimate of the maximal bandwidth - note that max_bw is > always increasing, so every 30 seconds we re-evaluate it. > To do that, my idea is to reset it, and force an increase in quality to > load the network, and re-estimate the bandwidth : > > max_bw /= 2.0; > appData.qualityLevel += 10; > > It's a hack, but I'm convinced that evidence-based empirical BWmax computation > is the only way to make this work - otherwise it's way too imprecise. > My testing shows pretty good results. > > 2-Special case: specifying BWmax > > As per my first mail in this thread, I have added -autoqual-bw as a > commandline > option, used to set up a known, constant BWmax. Useful when the user knows the > bandwidth of the network, or to limit how much BW is consumed by the viewer. > If you have a single viz node with a gigabit link and a standard provision of > 10 users, you can use -autoqual-bw 12000 to limit each viewer to 12MB/s. > > 3-If you want to force reevaluating BWmax > > Request an ALR or manually set a high quality for JPEG, this will overload the > network and speed up BWmax's increase > > 4-What do do with BWmax > > Using this to apply the algorithm we talked about earlier (target BW = 90%-95% > BWmax) seems, on my limited testing done in my specific environment, to work > relatively well in picking the quality. > > See at the bottom of this message for an output example. > > 5-When does this apply > > This is useful when : > > - the network is the bottleneck (this is the case on my 100Mbit/s connection > with 2*2GHz CPUs on each side, by the way). > - if you expect BWmax to vary during runtime because of e.g. congestion > (otherwise use -autoqual-bw) > - on a low latency network (I haven't tested this on a high latency network > yet, > but it will lead to underestimate BWmax and set a quality that is > too low, perhaps this isn't a big problem) > - at any rate, it can be disabled > > 6-Status > > This is merely a prototype that worked in a very, very limited test setup, but > I've been building understanding and experience with VNC those past weeks, and > I am more confident in this try than in the previous ones. > > Automatic quality adjustment is a very hairy issue, and I think a not-so-great > implementation is better than the current none-at-all state of TurboVNC, and > the static-and-inaccurate technique used by TigerVNC. > > What are your feeling about this? When I have done more testing I'll be able > to > show an actual patch (there is very little code, in fact). > > 7-Some output examples > > (Note that it changes the quality way too often, but that's easy to fix and > for > later!) > > 100Mbit/s > > Bandwidth 9511.44 kB/s (2.01 seconds) > Last BW sample is outdated, resetting max_bw to 0.000000 > quality set to 95, subsampling 2 > Bandwidth 11013.35 kB/s (2.02 seconds) > max bw 11013.351562 Link used at 100% > quality set to 85, subsampling 1 > Bandwidth 6865.89 kB/s (2.01 seconds) > max bw 11013.351562 Link used at 62% > quality set to 87, subsampling 1 > Bandwidth 7224.38 kB/s (2.00 seconds) > max bw 11013.351562 Link used at 65% > quality set to 89, subsampling 1 > Bandwidth 7840.85 kB/s (2.00 seconds) > max bw 11013.351562 Link used at 71% > quality set to 91, subsampling 1 > Bandwidth 8603.68 kB/s (2.01 seconds) > max bw 11013.351562 Link used at 78% > quality set to 93, subsampling 1 > Bandwidth 9719.05 kB/s (2.00 seconds) > max bw 11013.351562 Link used at 88% > quality set to 95, subsampling 1 > Bandwidth 11172.56 kB/s (2.02 seconds) > max bw 11172.559570 Link used at 100% > quality set to 85, subsampling 1 > Bandwidth 6854.07 kB/s (2.01 seconds) > max bw 11172.559570 Link used at 61% > quality set to 87, subsampling 1 > Bandwidth 7208.06 kB/s (2.01 seconds) > max bw 11172.559570 Link used at 64% > quality set to 89, subsampling 1 > Bandwidth 7841.28 kB/s (2.00 seconds) > max bw 11172.559570 Link used at 70% > quality set to 91, subsampling 1 > Bandwidth 8647.85 kB/s (2.00 seconds) > max bw 11172.559570 Link used at 77% > quality set to 93, subsampling 1 > Bandwidth 9721.50 kB/s (2.00 seconds) > max bw 11172.559570 Link used at 87% > quality set to 95, subsampling 1 > > 4500kB/s (simulated by trickle -d 4500, so kind of inaccurate) > > Bandwidth 3530.50 kB/s (2.02 seconds) > Last BW sample is outdated, resetting max_bw to 0.000000 > quality set to 95, subsampling 2 > Bandwidth 3185.69 kB/s (2.05 seconds) > max bw 3185.692383 Link used at 100% > quality set to 85, subsampling 1 > Bandwidth 3768.04 kB/s (2.01 seconds) > max bw 3768.036133 Link used at 100% > quality set to 75, subsampling 1 > Bandwidth 3595.85 kB/s (2.03 seconds) > max bw 3768.036133 Link used at 95% > quality set to 65, subsampling 1 > Bandwidth 3569.32 kB/s (2.01 seconds) > max bw 3768.036133 Link used at 94% > Bandwidth 3626.69 kB/s (2.03 seconds) > max bw 3768.036133 Link used at 96% > quality set to 55, subsampling 1 > Bandwidth 3669.86 kB/s (2.00 seconds) > max bw 3768.036133 Link used at 97% > quality set to 45, subsampling 1 > Bandwidth 3599.94 kB/s (2.00 seconds) > max bw 3768.036133 Link used at 95% > quality set to 35, subsampling 1 > Bandwidth 2776.69 kB/s (2.03 seconds) > max bw 3768.036133 Link used at 73% > quality set to 37, subsampling 1 > Bandwidth 2873.82 kB/s (2.01 seconds) > max bw 3768.036133 Link used at 76% > quality set to 39, subsampling 1 > Bandwidth 2978.42 kB/s (2.01 seconds) > max bw 3768.036133 Link used at 79% > quality set to 41, subsampling 1 > Bandwidth 3018.48 kB/s (2.03 seconds) > max bw 3768.036133 Link used at 80% > quality set to 43, subsampling 1 > Bandwidth 3145.13 kB/s (2.01 seconds) > max bw 3768.036133 Link used at 83% > quality set to 45, subsampling 1 > Bandwidth 3214.31 kB/s (2.00 seconds) > max bw 3768.036133 Link used at 85% > quality set to 47, subsampling 1 > Bandwidth 3310.21 kB/s (2.02 seconds) > max bw 3768.036133 Link used at 87% > quality set to 49, subsampling 1 > Bandwidth 3370.77 kB/s (2.02 seconds) > Last BW sample is outdated, resetting max_bw to 1884.018066 > quality set to 59, subsampling 1 > Bandwidth 3729.38 kB/s (2.01 seconds) > max bw 3729.376709 Link used at 100% > quality set to 49, subsampling 1 > Bandwidth 3415.06 kB/s (2.02 seconds) > max bw 3729.376709 Link used at 91% > Bandwidth 3386.41 kB/s (2.04 seconds) > max bw 3729.376709 Link used at 90% > > ------------------------------------------------------------------------------ Achieve unprecedented app performance and reliability What every C/C++ and Fortran developer should know. Learn how Intel has extended the reach of its next-generation tools to help boost performance applications - inlcuding clusters. http://p.sf.net/sfu/intel-dev2devmay _______________________________________________ VirtualGL-Devel mailing list VirtualGL-Devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/virtualgl-devel
