Posts by DigiK-oz

1) Message boards : Number crunching : Discussion of the new credit system (Message 24704)
Posted 24 Aug 2006 by DigiK-oz
Post:



Maybe, maybe not. I would have to do some longterm calculations on this, but point is that you will probably not gain much yourself. By the time you get the next WU from the same batch in, the effect of your previous claim will have diminished. Especially if the WU has had a lot of results reported already. Chances that some of your teammates gain anything are slim, too. And even if you do gain anything, fact is that other members/teams will have a higher chance of benefitting from you, unless you're in a team that is over half the active participants. So, you might be winning in actual points, but other teams will be winning more. So you're actually helping your opponents more than you help yourself.

Having said that, there's a loophole in that when someone reports their work in large batches (like 10 WU's). In that case, 9 out of the 10 would benefit (assuming that no ther WU's are sent by other members inbetween yours). Still, much more than 9 following your contribution would benefit....

Besides, your reasoning goes for the current system as well - and more so, because you gain immediately no matter what the effect on your next or teammate's WU.


G'day again LosAlcoholicos~Sloom

"Maybe, maybe not" - that uncertaintity would be enough for some individuals and/or teams to do what I'm hypothesising. That afterall, has been amongst the reasons given over the last weeks/months as to why optimised clients are used - "everybody else is using them" "team ?? are using them so why shouldn't we" "os ?? users are using them so why shouldn't we" "my cpu is disadvantaged compared to your cpu, so I'm using them".

"by the time you get your next wu from the same batch in" - what about mutiple hosts - same user multiple hosts? Never see the same wu type across multiple hosts?

For every cynical & devious idea I'm thinking up, I bet that there are ten more out there in the wild. Prove me wrong - and I say that without cynicism, because I hope that I am wrong. Only time will tell.



I will not prove you wrong, or try to counter your devious ideas. The suggestion I made just takes one extra reason OUT of overclaiming. It doesn't prevent every single cheat you can come up with.

In the current system, overclaiming gives you extra credit, and MIGHT make a difference on your next WU or your teammate's. But YOU are the main beneficiary.

In my suggestion, the first reason is gone. What remains is you MIGHT get more credits on your next WU or your teammate's.

So, I still think that it could be implemented, and probably very easy :

Cuurent system : Calculate a new rolling average, then assign average points

My idea : assign averige points, then calculate new rolling average.
2) Message boards : Number crunching : Discussion of the new credit system (Message 24692)
Posted 24 Aug 2006 by DigiK-oz
Post:


Looking at your post, there might be a very easy solution : people get the running average, BEFORE their claim is taken into account. This would make it useless to overclaim, in fact it would NOT give you more credits, but it WOULD give more credits to those that report WU's after yours. How's that for discouraging optimized clients? You would only benefit from overclaiming if you were actually the first to report a certain WU (what are the odds?). In all other cases, overclaiming would give more credits to everyone behind you - but not yourself, thus you would not get higher in the ranking - the opposite would be true.


Example : first guy reports, claims and gets 100 credits. You report, claiming 300, getting 100, setting the running average to 200. Third guy reports, claims 100, and gets 200, running average goes down to 166, etc....



G'day Los Alcohilocos~Sloom

If you overclaim you won't get the benefit on your first result, but you will on every subsequent result from the same type of wu. Sounds like a great incentive to me - not only can I doctor my subsequent credits but also those of all my teammates.

The fact that everybody else potentially benefits is not really a disincentive, if you and your teammates get your results before everyone else.

Potentially this means as a team you all set your runtime to the minimum, and your credit claims to the maximum.



Maybe, maybe not. I would have to do some longterm calculations on this, but point is that you will probably not gain much yourself. By the time you get the next WU from the same batch in, the effect of your previous claim will have diminished. Especially if the WU has had a lot of results reported already. Chances that some of your teammates gain anything are slim, too. And even if you do gain anything, fact is that other members/teams will have a higher chance of benefitting from you, unless you're in a team that is over half the active participants. So, you might be winning in actual points, but other teams will be winning more. So you're actually helping your opponents more than you help yourself.

Having said that, there's a loophole in that when someone reports their work in large batches (like 10 WU's). In that case, 9 out of the 10 would benefit (assuming that no ther WU's are sent by other members inbetween yours). Still, much more than 9 following your contribution would benefit....

Besides, your reasoning goes for the current system as well - and more so, because you gain immediately no matter what the effect on your next or teammate's WU.
3) Message boards : Number crunching : Discussion of the new credit system (Message 24675)
Posted 24 Aug 2006 by DigiK-oz
Post:
First impressions of the new credit system.

It does nothing to address the use of optimised clients, the use of optimised clients will in fact be encouraged. The individual user may not get full benefit of their claim, but enough users reporting optimised results will heavily skew the granted credits. This promotes a "safety in numbers" type attitude - meaning it's harder to spot the overclaiming hosts.

By this I mean there is nothing to stop someone modifying their xml files such that they have benchmarks 1000 times greater than would actually be calculated by the standard client. How would that affect their results, if they happened to report first they would be granted their astronomical claimed credit. If they didn't report first, then they wouldn't immediately get any advantage from their claimed credit, however, that claimed credit would increase the total_claimed and have an impact on subsequent reporters by increasing the average. Every subsequent granted credit would be radically greater than it should be (what you may miss out on in your first returned result you would be granted in your second and subsequent results). It would only take a relatively small number of these rogue hosts to seriously skew the granted credits (upwards of course).

Now we get to the really insidious part. It's much harder to track down these rogue hosts, why? because of the obfuscation inherent in the system of averaging the claimed credits - all of a sudden everyone is granted more - but you can't track the culprits (unless Rosetta is prepared to, and nulify the falsified claims).

What does this mean - rampant inflation of credits.

I hope I am wrong, but on first reading of this credit system, it will actually encourage many people to claim whatever they want with impunity. They will not necessarily initially get the benefit of what they claim but they will get the benefit of every previous overclaimer.

If this occurs then the credit system will not actually reflect work done, it will reflect the ever increasing falsification of benchmarks.

Why will this happen - because no one has said that it shouldn't (therefore it must be allowed).


Looking at your post, there might be a very easy solution : people get the running average, BEFORE their claim is taken into account. This would make it useless to overclaim, in fact it would NOT give you more credits, but it WOULD give more credits to those that report WU's after yours. How's that for discouraging optimized clients? You would only benefit from overclaiming if you were actually the first to report a certain WU (what are the odds?). In all other cases, overclaiming would give more credits to everyone behind you - but not yourself, thus you would not get higher in the ranking - the opposite would be true.


Example : first guy reports, claims and gets 100 credits. You report, claiming 300, getting 100, setting the running average to 200. Third guy reports, claims 100, and gets 200, running average goes down to 166, etc....
4) Message boards : Number crunching : BOINC setup in corporate environment (Message 24548)
Posted 24 Aug 2006 by DigiK-oz
Post:
If you set the buffer large enough (i.e. 5 days or something) there probably will not be a real "queue" right at the start of the allowed time. The client might not even connect every day, let alone right at the start of the interval. Unless, off course, you set the client to return results immediately.
5) Message boards : Number crunching : A cynical viewpoint. (Message 23703)
Posted 20 Aug 2006 by DigiK-oz
Post:
Why don't you give it up guys. There is a firm standpoint from dr Baker saying there will be NO backdating. Not your favorite standpoint, but a standpoint nonetheless.

Opening several very similar topics about this in which you can agree won't change that, and will only clutter the board.
6) Message boards : Number crunching : Why are discussions about Rosetta taking place on other boards? (Message 23429)
Posted 19 Aug 2006 by DigiK-oz
Post:
This is the part of a post I like best. I hope it's a hoax post, and Dr Baker never said such a thing. Roughly 90% of his active hosts fit this description

PLEASE--try to think about the science and the benefits to humanity, and ignore the taunts from the little guys--they just don't matter--and I hope all this will blow over soon. You have been making a fantastic contribution, and I would hate to see this cause any disruption to our collabortive efforts together.

http://www.xtremesystems.org/forums/showpost.php?p=1662478&postcount=20



Maybe, just maybe, you read the whole thing wrong, and should have extended the bold font 2 or 3 words to the left. So, it's not the little guys that don't matter, but rather it is thier taunts that don't matter.

If you read it your way, it's a big insult to many contributors, which I am sure Dr Baker would NEVER do. If you read it the way I outlined, it's trying to convince a bunch of contributors not to leave just because of a flamewar.

Look at it this way : If you moved the bold font two words to the right, one could read that Dr Baker favours female participants. So please keep thing in perspective.
7) Message boards : Number crunching : No checkpoint in more than 1 hour - Largescale_large_fullatom... (Message 13895)
Posted 16 Apr 2006 by DigiK-oz
Post:
This is ridiculous. A simple home-cruncher, leaving his/her PC on for only a few hours per day, could get stuck on one of those WUs indefinitely!



It does sound like a rather stringent requirement.

Is this going to be the norm. from here on out?


The plan is to try to have it checkpoint more often, or at least try to dump everything to a memory image at program swaps. But there is a lot of data involved, and interrupting the model effects the model outcome adversely.


Well, NOT having checkpoints/memory images whatever will adversely affect the entire project, as the small home-crunchers are getting fed up with not getting any results in because the same WU restarts from scratch over and over again. Maybe there is a way to hand out these large WUs only to computers having a high RAC? Or only to people who have their run-time preferences set to 8 hours or something similar?
8) Message boards : Number crunching : No checkpoint in more than 1 hour - Largescale_large_fullatom... (Message 13882)
Posted 16 Apr 2006 by DigiK-oz
Post:
This is ridiculous. A simple home-cruncher, leaving his/her PC on for only a few hours per day, could get stuck on one of those WUs indefinitely!

9) Message boards : Number crunching : Xeon Performance (Message 8431)
Posted 5 Jan 2006 by DigiK-oz
Post:
Also, on multi-CPU XEON boxes (which hyperthreading is, sort of), the benchmarks are run simultaneously. Memory-bandwidth seems to become a problem there. To see if that is your problem, set your preferences to (temporarily) use only one cpu on multi-cpu machines, then rerun the benchmark. If it turns out significantly higher, memory may be constraining it. Why this does not hold true for everybody, dunno. Older XEONs? Worse memory? Worse motherboards? Could all very well be.
10) Message boards : Rosetta@home Science : does Rosetta/BOINC support a 'networked install' (Message 3217)
Posted 14 Nov 2005 by DigiK-oz
Post:
Thanks for the quick replies. I will try to have the clients reconnect to their previous share after reboot. However, clients come and go in this setup, so occasionally the need may arise (or some stupid error on my part may cause it) to have a different client connect to some directory. I don't want to end up with a bunch of directories that "were used once, and might be again in the near or distant future". What I would like to do in the case of some client being removed/replaced or whatever, is reuse that client's directory.

Well, from your replies I read that this shouldn't pose a real problem, but that a re-use by the same client would be preferred if possible.

Ok, I'll see if I can get it setup that way and take it from there.
11) Message boards : Rosetta@home Science : does Rosetta/BOINC support a 'networked install' (Message 3210)
Posted 14 Nov 2005 by DigiK-oz
Post:
Thanks for the suggestion. A ramdrive isn't ideal, since that approach would still be an install, leaving entries in control panel etc even when the ramdrive is long gone. Not to mention stalled jobs after reboot and who knows what else.

As for the bunch of copies of the install directory and sharing those, one for each client, that sounds promising. I already tried sharing a directory and running multiple clients from that single directory, but that failed miserably, with error messages, project resets and more issues. Having a seperate directory for each client is a little more of a hassle, but since that should be a one-time setup (maybe again when and if more clients are added) that's ok.

One more question about this : should every client, after a reboot for instance, reconnect to it's own directory (the one it used before)? Or would it be ok if it connected to some other directory (continuing the work some other client started), as long as there is always only one client connected to any single directory?

I'll be testing this somewhere later this week. Thanks again!
12) Message boards : Rosetta@home Science : does Rosetta/BOINC support a 'networked install' (Message 2744)
Posted 9 Nov 2005 by DigiK-oz
Post:
I would like something like this as well. At work, I am allowed to run DC projects on some 50 or so machines, but if, and only if :

- There's no install on the machines
- It doesn't use an internet connection

Coming from find-a-drug, I just shared the find-a-drug directory on one machine (which WAS allowed to have an install and internet connection). So, this machine, called let's say "fadserver" would share it's find-a-drug directory with sharename "fad". On any machine I'd want to run find-a-drug on, i'd type a single command :

\fadserverfadloader.exe

And the machine would start crunching, getting it's work from, and returning the result to, the machine called fadserver which took care of downloading/uploading work.

Any suggestions for setting up BOINC in this fashion? Using third-party software maybe? Or hints on writing a program which would make this possible? Any thoughts on the matter would be appreciated, and might help me contribute way more power...






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