Beta testing - Calibrated Boinc core

[Honza]

How about CPDN BOINC 5.3.11

[trux]

I didn't look yet what is the latest version available on the developer server, but I'll take the version of the last checkpoint. If it is 5.3.11, it will be 5.3.11, but I'll not use any project specific clients.

[trux]

Yes, the last available source code release on the Boinc server is the 5.3.11

[Mr.Pernod]

if there are no changes to the calibration-mechanism, I will stick to the 5.3.8.tx29 untill at least wednesday, I need the result-history to evaluate the calibration.

[trux]

No, the calibration mechanism did not change since a while. There is just the fix of the partial project name ambiguity, and fixes of other features or reporting.

[trux]

New release, 5.3.11.tx31 is online. I synchronized the source code with the official 5.3.11. I do not see any significant changes there. The only one that could be interesting (if it works), is some additional handling of HTTP responses - that may help avoiding the dropping of too many HTTP connections when the server is overloaded and dropping some TCP/IP packets. With the extension of HTTP handling, the client may be better in completing downloads/uploads even with some packets dropped.

On my side, there are changes I already announced: partial project names in calibration and priority settings are no more possible, and should cause no conflicts at projects with overlapping names (i.e. SETI@Home vs. SETI@Home Beta Test). Calibrating log message at the very first WU after the client restart should now show the final values, and at priority_projects, backup projects should request 1 second of work only _after_ the main project fails.

There is a new configuration option in this release: <delete_overdue>1</delete_overdue> for deleting workunits past their deadline. Currently only expired unfinished WU's of all projects are being automatically deleted. In future, other values than 1 may be used for deleting also WU's with estimated completion time past the deadline. It will be little bit more risky, and in some cases (changed benchmark, incorrect time estimation, wrong WU info from the server,...) it could even cause deleting of freshly downloaded units, so for now, I am keeping it just simple and delete only already expired WU's.

CAUTION: I did not test this feature at all yet! It may cause damage by deleting valid units, though I certainly hope it will not be the case.

[trux]

OK, I do not really have a lot of feedback, what can either mean that people simply do not care, or that there are no problems and it works as expected. I am thinking about releasing it today publicly, so if you have anything to report before I do it, please do it now. Also, I welcome any reports not only about bugs, but about your satisfaction with the calibrating both at S@H and at other projects.

[Julian Ellis]

good calibration of credits on Seti, Predictor and Simap all coming in as they should I have never had this before even on non-optimised apps and clients I use P4 3.2Ghz hyperthreaded (I think the hyperthreading caused even official apps and clients to claim only half what they should for me in the past) I have 14 projects attatched with Seti set for priority, I suspended seti to let other projects download so will let you know about them when they finish but that might take a while.

Thanks for all your hard work I am sure that most of the people here appreciate it, and I look forward to seeing what you do in the future

[Mr.Pernod]

I was hoping for a bit more time to evaluate the calibration, but if you want to release it, go ahead.
From what little history I have been able to build up under 5.3.8.tx29 for several project, the calibration seems to be within 15~20% of what the claimed credit should be when using a stock BOINC client.
calibration for "all" is working, but I still have to check for SETI Beta, the two results have I returned since installing these clients (5.3.8.tx29/5.3.11.tx30) did not have the extra info injected in the stderr.out on the website.

I have not really tested the latest 5.3.11 releases, except the tx30 which I run on the SETI Beta machine.
priority_projects seem to work without confusing "SETI@Home"and "SETI@Home Beta Test",
but if I remember correctly BoincManager 5.2.13 is not showing any messages when NNW is engaged/disengaged, it seems to work nonetheless.

[trux]

S@H Enhanced uses different crediting system - it does not use time measuirng and benchmarks at all, hence the calibration will have no impact on the credits. The credits are much more stable with Enhanced and will stay so until a new Enhanced client with considerably different crunching algorithm is developped - in that case, some kind of calibrating may be needed again, but the current one will need some modifications to work with Flops oinstead of time & benchmarks

The coming S@H Enhanced is also the reason I want to release the calirating client asap, so that it can be useful at least shortly. Would be a pity if I wasted all the time for nothing.

[Mr.Pernod]

I probably got confused by the "Gfpops: 2.60 >> 11.96" in the calibration report in BOINCManager, never mind.

a quick rundown of the results on my systems:
cpu_affinity => works
return_results_immediately => works
priority_projects => works
calibrate_credit => works within +/- 15~20% of the stock client, depending on the project.

I have not tested the other options.

[trux]

Just a note about the calibration - the stock client does not necessarily claim the correct credit. You can easily see it if you have machines with different performance - slow machines will mostly claim higher credit than the fast ones, and that's definitely not the intended purpose of the benchmark measured crediting system, BOINC introduced. The benchmark is simply insufficient for the measurment, because it cannot simulate the complexity of the project calculation, hence the speed gain varies. This is why the calibrating is needed. I proposed server driven calibrating many month ago, but since it was not introduced, I decided to develop at least this one. I think that mostly it is even much more precise than just 15-20%, although certainly, server-driven calibration would be better.

[Mr.Pernod]

The 15~20% I found is from both checking random results and from finding machines with comparable or the same cpu.
If I had to make a chart with a median and a standard deviation (no way, don't even think about asking that!!), it would probably be a nice bell-shape with the extremes at both plus and minus 20% claimed credit, but the majority close to what other systems on average claim.

[EDIT]
are you going to release v5.3.11.tx31 or are you going to do another compile when you go public?
[/EDIT]

[trux]

The 15~20% I found is from both checking random results and from finding machines with comparable or the same cpu.

That's exactly what you do wrong. You would need to compare it with a reference machine (calculating at the speed of 1Gfpops and 1Giops).

are you going to release v5.3.11.tx31 or are you going to do another compile when you go public?

v5.3.11.tx31 is currently the last one.

[Mr.Pernod]

The 15~20% I found is from both checking random results and from finding machines with comparable or the same cpu.

That's exactly what you do wrong. You would need to compare it with a reference machine (calculating at the speed of 1Gfpops and 1Giops).

I see.
I was looking at the numbers other users see just before they start screaming "cheater, cheater"
If you happen to know which processor(s) they use as reference, I can try to locate one and compare numbers, no problem.


Note about the NNW with BoincManager 5.2.13 and core client 5.3.11, BoincManager does not display a message in the messagetab when toggling the function, but it does work.