New Boinc Windows core client released - 5.2.1
11 Oct 2005 19:04:45 UTC
Topic 190001
(moderation:
Any one testing the 5.x.x Boinc clients?
The new Boinc core client for Windows is available, the even middle integer indicating a release version (thanks again, Jord for the nomenclature explanation :) ), but curiously, it is still listed as a development version, large red message, may be unstable, use only for testing.
I've been using the 5.1.8 alpha for the past 2 days, now processing the 10th WU under it, with no problems.
A quirk(or design feature?...maybe to avert overloaded work queues) in 5.1.8 *overestimates* time to completion, but only during processing, decrementing TTC very slowly through the first hour or so, until the total time is more than an hour greater than it should be. During the mid-portion of the WU, TTC decrements equally to the incrementation of CPU time. When the WU is approx 1.5 hr from *actual* completion, displayed TTC begins to accelerate decremtation, increasing toward a factor of 2x processing time, and converging upon *real* time remaining as progress approaches 99%.
Note - these times and figures apply to my ~5hr11min WUs.
"To completion" estimates for queued WUs is *very accurate*, and seems to adjust according to recent completions, probably some kind of adaptive strategy at work there. Anyway, this aspect is much improved over 4.x clients, for which we may be grateful. Good work!
Kind regards to the dev team.
microcraft
"The arc of history is long, but it bends toward justice" - MLK
New Boinc Windows core client released - 5.2.1
)
Assuming no major bugs are found today it should be moved to the recommended version tomorrow. We actually found one in 5.2.0 or it would have been out today.
BOINC WIKI
BOINCing since 2002/12/8
RE: Assuming no major bugs
)
Is that the ZoneAlarm AV problem? How is the workaround coming along?
microcraft
"The arc of history is long, but it bends toward justice" - MLK
RE: RE: Assuming no
)
The big bug was in new installs it would crash.
I think there was a fix in 5.2.0 for the ZA problem.
BOINC WIKI
BOINCing since 2002/12/8
At conclusion of Boinc 5.1.8
)
At conclusion of Boinc 5.1.8 trial: Crunched 13WUs, 9 validated & granted credit, 4 pending.
No problems; only oddities noted in earlier post.
Unable to test graphics or screensaver, as I long ago deleted .scr file. Other than that, all buttons, bells, & whistles function as I expect.
~12 hrs ago, successfully transitioned to Boinc 5.2.1
1st WU completed, pending. 2nd WU nearing completion.
Remarks re: 5.1.8 apply to 5.2.1 also. No problems running alongside my NOD32 AV.
System: AMD AthlonXP Barton, 512MB L2 cache, Windows XP Home sp2, 512 MB RAM
microcraft
"The arc of history is long, but it bends toward justice" - MLK
downloaded boinc 5.2.1
)
downloaded boinc 5.2.1 (windows), attached to einstein, and have uploaded 2 wus so far. everything is working well. completion time is very close to actual crunching time.
RE: downloaded boinc 5.2.1
)
Great, Devn! Keep us posted. If you don't mind, list important system specs...Operating System, CPU, Einstein version, etc.
The bug (I said "quirk, or design feature" in my first post) is still there. Midway through a WU, the "To completion" time has become progressively inflated (by more than 50%), to the extent that total time (CPU time + To completion) has become overestimated by about 28%, progressively reaching that point and then gradually decreasing, approaching actual total, through the 2nd half of processing. For example - my PC crunches a WU in about 5hrs, 10min. So at 50% completion, CPU time reads 2 hrs 36min, correctly. But the "To completion" reads about 4hrs!
It is unlikely, but possible that this behavior was purposely built in, but it is questionable. What's the use of displaying "To completion" if it's wholly inaccurate? I'm at the point where I call it a bug. Too bad, because otherwise I'd give 5.2.1 a big Thumbs UP.
microcraft
"The arc of history is long, but it bends toward justice" - MLK
RE: The bug (I said
)
Amendment to my above quoted post:
I arrived at this conclusion with reluctance, because in every other way that I was able to test this release, it has performed superbly, ready for primetime, through 7 crunched WUs, 2 validated already. In retrospect, I would term this a "flaw", rather than a "bug". If there is a more appropriate place to report this, I welcome advice. If nothing else, I think that a Forum Moderator may wish to move or copy this post to the Bug Reports message board, where a member of the development team would be more likely to take note. Those folks have worked harder than I could know, and have my great respect and appreciation.
microcraft
"The arc of history is long, but it bends toward justice" - MLK
Michael, RE: If
)
Michael,
How about on the BOINC Messages Board at http://boinc.berkeley.edu/dev/
But before you go jumping there read this first:
Hi, Stick, glad to hear from
)
Hi, Stick, glad to hear from you.
Thank you for the pointer to the Boinc dev board.
re: the "To completion" estimates, from what I can tell, there are 2 distinct changes from the approved client.
#1 - for *queued* WUs, the estimate is far more accurate (in 5.x) than the 4.x, which gave inflated "To completion" times.
#2 - for *in progress* WUs, the "flaw" I noted distorts the "To completion" estimates throughout processing (in 5.x), whereas the 4.x core clients I've used have corrected the inflated "To completion" figures (see #1, above) very accurately and very quickly after processing begins, i.e., on the 4.x Boinc client, once a WU was 5 minutes or so in progress, you could set your watch by the to completion estimate and it would be dead on (excluding variances due to varying competing workloads during the time Boinc/Einstein was running), except for the last 1%, which took virtually no time at all.
Addit - The 2 things above regarding behavior under Boinc 4.x did not change when I tested the Einstein 0.18 beta app, or when I changed back, but only when I began testing the Boinc 5.x client, leading me to conclude that the 5.x client is responsible.
It should be noted that I run *exclusively* Einstein, and so I cannot extrapolate my observation to include any other combinations of clients and science apps, or to Boinc in general, but limit it to Boinc/Einstein only.
microcraft
"The arc of history is long, but it bends toward justice" - MLK
Michael, Again, comparing
)
Michael,
Again, comparing my "To completion" experience with BOINC 4.45 to your 5.2.1 experience raises some questions:
[qoute]#1 - for *queued* WUs, the estimate is far more accurate (in 5.x) than the 4.x, which gave inflated "To completion" times.[/qoute]
My 4.45/0.18 queued estimates are usually low - about 70% of what it actually takes to complete. However, queued Seti WU's are typically estimated at about 3 times what it actually takes.
[qoute]#2 - for *in progress* WUs, the "flaw" I noted distorts the "To completion" estimates throughout processing (in 5.x), whereas the 4.x core clients I've used have corrected the inflated "To completion" figures (see #1, above) very accurately and very quickly after processing begins, i.e., on the 4.x Boinc client, once a WU was 5 minutes or so in progress, you could set your watch by the to completion estimate and it would be dead on (excluding variances due to varying competing workloads during the time Boinc/Einstein was running), except for the last 1%, which took virtually no time at all.[/qoute]
My 4.45/0.18 experience generaly agrees with what you described but Seti is quite different. In progress Seti WU estimates, start out by immediately dropping from the very overinflated queued numbers to near zero. Then they rises in parallel with elapsed CPU time numbers until the unit reaches about 50% completion. Then, after reaching the 50% mark, Seti "To completion" estimates become very accurate.
My point is this: Although you may think the new BOINC has some differences from the 4 series with regard to "To completion" issues, I don't see how BOINC can account for the very different behaviors I have observed between Seti and Einstein WU's. That is, it seems to me that the place to fix these estimating problems would be in the Science Apps not in BOINC.
Stick