Posts by Webmaster Yoda

1) Message boards : Number crunching : Supported OS (Message 9814)
Posted 25 Jan 2006 by Profile Webmaster Yoda
Post:
But what with the rosetta-client?
For which OS is there support?
Also for *BSD, Sparc, ...?


According to http://boinc.bakerlab.org/rosetta/apps.php Rosetta is available for Linux/x86, Windows/x86 and Mac OS X

AFAIK the source code is not available for compiling elsewhere (unlike with SETI).
2) Message boards : Number crunching : intel3.2dc versus intel 3.8 (Message 9500)
Posted 21 Jan 2006 by Profile Webmaster Yoda
Post:
NO Dualcore AMD advise, mmmm


If you look through the replies (3rd and 4th) you will see that I did mention the Athlon X2 4400, but James gets discount from a supplier who only does Intel CPUs. Hence we're talking about which Intel, rather than which brand CPU.
3) Message boards : Number crunching : intel3.2dc versus intel 3.8 (Message 9433)
Posted 20 Jan 2006 by Profile Webmaster Yoda
Post:
The P4 3.8 has HT allowing it to run 2 units at a time, a bit faster than the duel core 3.2


Sorry, I'm with Paul on this one, based on my experience with the 3.4GHz version.

A P4 with HT will run two WU at maybe 70, maybe 75% efficienty each (i.e. about 2.66-2.85GHz each on a 3.8GHz CPU, sharing the cache) The 3.2GHz dual core on the other hand will run each WU on a dedicated 3.2GHz processor with its own cache.

Oh, and there is a 3.4GHZ Dual Core now, with 2x2MB L2 Cache. The CPU is not much more expensive than the single core 3.8GHZ from what I have seen.
4) Message boards : Number crunching : intel3.2dc versus intel 3.8 (Message 9328)
Posted 19 Jan 2006 by Profile Webmaster Yoda
Post:
Basically, I can get a 3.2ghz DC for 1k under the dell list price for both home and business users. The 3.8ghz is about 40 more to add.


OK, go for the dual core and use the money you save to add a decent CPU heatsink and fan (like a Zalman CNPS9500LED or similar :-)
5) Message boards : Number crunching : intel3.2dc versus intel 3.8 (Message 9237)
Posted 18 Jan 2006 by Profile Webmaster Yoda
Post:
Anyone have the benchmarks for the 3.2ghz dual core and the 3.8 ghz? I'm going to buy a box and as the price is marginal I would like to know what I'd be losing by purchasing the 3.2ghz.


Can't help you with benchmarks, but a dual core 3.2GHz will probably do more crunching than a single core 3.8GHz (even with HT on).

Hard to compare with the variable length of Rosetta WUs but looking at my own 3.4GHz P4 with HT, it has approximately the throughput of two 2.4GHz P4's when running SETI. I'd expect the 3.8GHz to have a little less throughput than a dual core 2.8GHz.

Having said all that, I'd look at an Athlon X2 instead. Lower power consumption, less heat (even if overclocked). Don't know what prices you're looking at, but an Athlon X2 4400+ is likely cheaper than either of the Pentiums and should easily outpace them (even at stock speed).

[EDIT]Note: A motherboard for the Athlon will probably also be cheaper[/EDIT]
6) Message boards : Number crunching : Leave in Memory? (Message 9118)
Posted 16 Jan 2006 by Profile Webmaster Yoda
Post:
I don't leave mine in memory and I run Rosetta just fine.


If just fine means it completes SOME work units OK, I guess you're right.

Your Athlon X2 (host id 765) has nearly as many WUs with client errors as it has successes - I wonder how many of those are due to you not leaving them in memory.
7) Message boards : Number crunching : Stats of 1 machine? (Message 8952)
Posted 13 Jan 2006 by Profile Webmaster Yoda
Post:
I really like to see the output of every computer, so if they fail i can see it within one day....


You could have a look at BOINCstats, in particular this page for your computers.

It does not show individual work units but it does give a good overview of how much work each of your computers has done in the last 24 hours, week and month (and if you click on the stats for a single computer you get even more info)
8) Message boards : Number crunching : AMD vs Intel (Message 8756)
Posted 11 Jan 2006 by Profile Webmaster Yoda
Post:
Does anyone have a list of all the BOINC projects and the processors they 'prefer'? As in AMD vs Intel?


I'd be interested in that as well. From my own recollection though:

  • Rosetta was fastest on Athlon (XP/64)
  • SETI (with optimised application) works best on Intel with SSE2 or SSE3 instruction set, and Athlon 64 with SSE3:

    • 1.7GHz Celeron (SSE2), Linux: avg 1.5 hours
    • 2.3GHz Athlon XP 3000+ (SSE), WinXP: avg 2 hours+
    • 2.4GHz Pentium4 (SSE2), WinXP: avg a little under 1 hour
    • 2.6GHz Athlon 64 3700+ (SSE3), WinXP64: avg 40 minutes
    • 3.4GHz Pentium 4, HT (SSE3), Win2003: avg 55 minutes for 2 WU


  • The old (pre Albert app) Einstein worked best on Athlons (XP/64). From memory, the Celeron took 16 hours, the P4/2.4 took 10 hours, the Athlon XP took 5.5 hours, the Athlon 3700+ took less than 5 hours and the Pentium 4/3.4 with HT took 11-12 hours for two.
  • Predictor also worked best on Athlon XP/64
  • CPDN likes fast Pentiums (even with HT on) better than Athlon


But your mileage may vary :-)

9) Message boards : Number crunching : Only one computer per wu? no checks? (Message 8745)
Posted 10 Jan 2006 by Profile Webmaster Yoda
Post:
For some reason 11 crossed my mind after I had posted; 10 error + 1 quorum. - the fun of programming :)


For i=0 to 10.... I think that's what's happening here. The very first WU sent out is number 0, not number 1 (have a look at the last number of the WU name on a work unit in your cache) I'm no programmer, but javascript arrays start with element 0 as well.


10) Message boards : Number crunching : Please abort WUs with (Message 8717)
Posted 10 Jan 2006 by Profile Webmaster Yoda
Post:
Well yes, but at least on my machine the 'time to completion' only decreases when the percentages change (say, from 20% to 30%), but it increases during periods where the percent value stays constant.


Makes sense to me. If after 30 minutes it's at 10%, the estimated time to completion would be 10 times the time already taken (i.e. 5 hours) and keep climbing until it hits 20% (If I spend X minutes to do Y% how long will it take to do 100%?) It's further complicated by the variable length of the steps.

I have suspended crunching for Rosetta for now, but never took much notice of the estimated time to completion as it was basically meaningless.
11) Message boards : Number crunching : is 512MB RAM really needed? (Message 8277)
Posted 3 Jan 2006 by Profile Webmaster Yoda
Post:
But the work shouldn't be completely lost--if you have already written out five structures when you turn off your computer off, the next time you turn it on it should only have to make five more (I need to check on this).


I'm aware of that. The issue is that, especially on the longer WUs, each of these structures can take a long time to do, so a lot of calculations are wasted when computers are switched off.

It is midnight here and I'm about to switch the computers off. One of them is showing 1% complete after 26 minutes crunching (on an Athlon XP 3000+, a reasonably quick machine). Unless it advances to 10% in the next few minutes, that half hour of crunching will be wasted as we go back to zero in the morning.
With a slow machine (say the minimum, 500MHz PC), the steps could be hours apart.

Most other projects are able to show more even progress and save more often. I'd like to see Rosetta do that too, if at all possible.
12) Message boards : Number crunching : is 512MB RAM really needed? (Message 8244)
Posted 3 Jan 2006 by Profile Webmaster Yoda
Post:
We've been working on this, but we are running up against some hard limits that will be hard to overcome. Most of the jobs going out these days should be comfortably under 200 MB.


Maybe I'm over-simplifying and maybe there's no solution but I just get the feeling that the memory use and lack of CPU checkpoints and saves are related.

With some of the latest batch of work units taking 9 hours* on a 3.4GHz Pentium 4, there's a lot of crunching being lost when I shut down PCs at night, reboot them (for other, unrelated issues), etc. On slower machines it can be many hours between saves.

Maybe if Rosetta could checkpoint and save to disk more often it would not need to keep as much data in (virtual or real) RAM?

* NO_RANDOM_WTS_OR_FRAGS_1ogw_223_61_2 took just over 9 hours - I wouldn't like to run that on a 500MHz machine!
13) Message boards : Number crunching : 2 million a day (Message 7946)
Posted 30 Dec 2005 by Profile Webmaster Yoda
Post:
I would point out that there were two dates given in the same sentence; 1/4/06 and 1/15/06. If you want to make the first one April 1, that's fine... but what is the second one?!?


Probably something like ***UNHANDLED EXCEPTION****
Reason: Access Violation (0xc0000005) at address 0x00656358 read attempt to address 0x2A08CFE8

Sorry, couldn't resist :-)
14) Message boards : Number crunching : again on on computation error (Message 7893)
Posted 29 Dec 2005 by Profile Webmaster Yoda
Post:
Further to what I wrote below...

I run both SETI and Rosetta on most of my computers (currently on a 50/50 basis). I have "Leave applications in memory while preempted" set to "yes" and (apart from the recent bad batch) rarely have a problem.

I can suspend individual work units. I can suspend Rosetta while it's running a work unit. I can switch to SETI. Not a problem. But if I set the above to "no", I will get many crashes, guaranteed.
15) Message boards : Number crunching : again on on computation error (Message 7883)
Posted 29 Dec 2005 by Profile Webmaster Yoda
Post:
One question, is true that was write (don't remember where) that changing the preference of "processor usage", "Leave applications in memory while preempted" from No to Yes, can help to solve the problem?


The first three are clearly from the bad batch mentioned in other threads. Nothing you can do about those. The last one is not from that batch but without more details it's hard to tell what caused the error.

Changing the setting to "yes" would probably have prevented the error if it happened during benchmarks or switching to another application.
16) Message boards : Number crunching : RAC up & down (Message 7718)
Posted 27 Dec 2005 by Profile Webmaster Yoda
Post:
RAC shows who is currently contributing the most to the project rather than who has run the most calculations.


I can be useful, except it is an inaccurate figure. Someone can build up am RAC that's 10 times anyone else's on a project, then stop crunching. It will take time before that RAC drops to reflect what is "currently" happening. The opposite is also true - it takes a while to build up the RAC.

But I guess it depends on one's interpretation of "current" :-)
17) Message boards : Number crunching : Sometimes with Boincmanager time changes backwards. (Message 7657)
Posted 26 Dec 2005 by Profile Webmaster Yoda
Post:
I only run R@H and some of the machines have only got 256MB RAM, so don't want to slow them down too much.


256MB is below recommended specs but it should not be a problem in itself, particularly if the machine only runs Rosetta.

Would setting "leave apps in memory" to yes change something when I run only one application? Will I lose work done for instance?


I'd say the opposite is the case. With a setting of "no", any time Rosetta or the work unit is suspended you lose CPU time spent since the last save (which can be quite a while). The work unit can be suspended for a number of reasons, including "no" on the the first two settings in preferences, running benchmarks, or a manual suspend.

Another reason it may have gone backwards would be if BOINC was restarted or the computer rebooted during the 10 minutes you mentioned (in which case, obviously, none of the settings would make a difference)

If none of the above happened, I don't know what caused the clock to go backwards - maybe someone else has ideas.
18) Message boards : Number crunching : Sometimes with Boincmanager time changes backwards. (Message 7653)
Posted 26 Dec 2005 by Profile Webmaster Yoda
Post:
Sometimes when I check Boincmanager again after a few minutes, I see the passed time going backwards.


It sounds like you may have the following setting in your preferences:
Leave applications in memory while preempted? no
Switch between applications every 30 minutes (or maybe less)

Rosetta has a known bug (documented in several threads on these boards) which is usually fixed by changing that first setting to yes.

EDIT: Rytis, you beat me to it :-)
19) Message boards : Number crunching : Please abort WUs with (Message 7643)
Posted 26 Dec 2005 by Profile Webmaster Yoda
Post:
If everybody else takes 4 days to get to this one, we'll be into February. So yes, the staff needs to "kill" these, can't rely on them being gone by the time they get back. That's the question I was trying to answer, it's just not the answer I hoped for. :-(


Yeah, the 10 errors alllowed per WU is going to keep these (and the other bad WUs) circulating for some time.

If this can't be fixed in the scheduler, perhaps it can be fixed by deleting (or renaming) the directories on the project's server that these bad WUs are stored in.

It would result in download errors but it would reduce bandwidth usage, stop anyone crunching them and get them flushed out of the system quickly (assuming download errors add to the error count on the WU)
20) Message boards : Number crunching : No Work (Message 7588)
Posted 25 Dec 2005 by Profile Webmaster Yoda
Post:
The Boinc Folder contains a total of 132 files in 40 folders, and uses a total of 34.7 MB.


Sure sounds like BOINC is confused. Have you tried Bill's suggestion (shown below)?

This is going to sound silly, but try it... change that to 60%, hit "update" on the website, go back to BOINC Manager, select Rosetta in the Projects tab and hit Update."


Next 20



©2021 University of Washington
https://www.bakerlab.org