Message boards : Number crunching : Re Toni_TAKE3
| Author | Message |
|---|---|
|
TylerChris Send message Joined: 12 Feb 10 Posts: 11 Credit: 50,020,466 RAC: 0 Level  Scientific publications    | |
|
Just completed one of these WUs, took around 2 hours. | |
ID: 23475 | Rating: 0 | rate:
 /   Reply
Quote
| |
 skgivenVolunteer moderator Volunteer tester  Send message Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level  Scientific publications    | |
|
So you ran one normal length task and jumped to the conclusion that GPUGrid has been knocked down the pecking order! | |
| ID: 23479 | Rating: 0 | rate:
/ Reply
Quote
| |
|
I thought I read somewhere that the ToniTake3 in the long queue were | |
| ID: 23481 | Rating: 0 | rate:
/ Reply
Quote
| |
|
skgiven Volunteer moderator Volunteer tester Send message Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level Scientific publications | |
|
Yeah, that's part of the plan, if any more are released as Long tasks (whatever the name). | |
| ID: 23483 | Rating: 0 | rate:
/ Reply
Quote
| |
|
Toni Volunteer moderator Project administrator Project developer Project tester Project scientist Send message Joined: 9 Dec 08 Posts: 1006 Credit: 5,068,599 RAC: 0 Level  Scientific publications  | |
|
TAKE3, being in the long queue, should have been double credits. That was my mistake. I did not cancel them because there are not many of them. | |
| ID: 23484 | Rating: 0 | rate:
/ Reply
Quote
| |
|
Toni Volunteer moderator Project administrator Project developer Project tester Project scientist Send message Joined: 9 Dec 08 Posts: 1006 Credit: 5,068,599 RAC: 0 Level Scientific publications | |
|
I've made an emergency fix so that TAKE3 jobs will give 10000 credits + bonuses each, but this will apply only the the newly-created workunits. | |
| ID: 23491 | Rating: 0 | rate:
/ Reply
Quote
| |
[PUGLIA] kidkidkid3 Send message Joined: 23 Feb 11 Posts: 81 Credit: 954,353,044 RAC: 218,738 Level  Scientific publications | |
|
Hi all, | |
| ID: 23493 | Rating: 0 | rate:
/ Reply
Quote
| |
|
skgiven Volunteer moderator Volunteer tester Send message Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level Scientific publications | |
|
Welcome to the Forum. | |
| ID: 23494 | Rating: 0 | rate:
/ Reply
Quote
| |
|
TylerChris Send message Joined: 12 Feb 10 Posts: 11 Credit: 50,020,466 RAC: 0 Level Scientific publications | |
|
Hi skgiven So you ran one normal length task and jumped to the conclusion that GPUGrid has been knocked down the pecking order! Err no Used that task as an example to show the discrepancy as have not crunched any ACEMD2 tasks for a while,so had nothing to link to. Your 470s crunching at donate @home here To summarise they took around 3700 secs for 6250 credits. now take a look at the same gpus crunching ACEMD2 tasks here. Think the comparison is pretty clear.Around twice as many credits are granted for filling the coffers as is for the science. You wrote; I have already worked out that the credit CC2.0 cards get by running Long tasks here is the same or slightly more than D@H (often 10 to 20% more, depending on the task), at least for well optimized systems Yes the Long tasks, a good amount of crunching here is on the ACEMD2 tasks that do not carry a 50% bonus .Take the bonus away then what do you have? do the donate at home have a 24 hour deadline with less credits granted if not reached? You wrote; You are using W7, which is at least 11% slower than XP here. Even so, running Long tasks here you would on average get ~10% more credit, not half the credit! What has windows got to do with the argument? yes windows 7 is a little slower but does not change the maths much. Have withdrawn the card from the project because I think the projects values are wrong, in my opinion the science should be valued more highly than bitcoin mining from it's volunteers. The credit system as it stands as I see it is,crunch ACEMD2 we will award you a few credits. But come over to donate@home and stick money in the coffers and we will give you at least double credits. Hope this makes my position clearer. | |
| ID: 23502 | Rating: 0 | rate:
/ Reply
Quote
| |
|
skgiven Volunteer moderator Volunteer tester Send message Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level Scientific publications | |
|
Suggestions were already made to increase the credit for normal tasks where they are returned very earlier. The was primarily to allow people with high end cards to run some normal tasks (badge system related) without seeing a massive dip in credit RAC. This project is different and weight is put behind the fast return of long results as their actual value to the project is much greater than the slow return of the short tasks. The credit here accurately reflects the needs of this project, and the input of fast turnover vs slow. The credit system here has been modified and adapted over the years, accounting for things such as application performance increases. It does not take into account other unrelated and uncomparable credit systems, especially those for projects that turned up after this one. | |
| ID: 23503 | Rating: 0 | rate:
/ Reply
Quote
| |
SMTB1963   Send message Joined: 27 Jun 10 Posts: 38 Credit: 524,420,921 RAC: 0 Level  Scientific publications | |
|
skgiven writes: The credit at D@H is 10 or 20% lower for top NVidia card there than it is here if running the Long tasks ATI/AMD cards are about 3 times as fast at D@H so are much better suited to that project. It's astonishing to me that anyone could look at these two statements and fail to see a problem. You can argue all you want about TylerChris' specific example, but IMO he has raised a valid point. @TylerChris: I hope you will reconsider withdrawing your support...things are still at a very early stage with D@H, and my guess is that policies regarding these matters are still being formulated. Despite skgiven's somewhat argumentative tone, the leadership here has demonstrated they are listening to volunteer concerns - so no need to do anything rash ATM. : ) We are also around a month from seeing GF600 cards, which may or may not improve upon their implementation of the calculation that allows ATI to card perform better. So what? If Kepler changes the bitcoin performance equation when it comes out, fine. In the meantime, it is ridiculous to purpose ANY nVidia card for bitcoin mining. You know it, I know it, GDF knows it, and every noob miner on the bitcoin network knows it. While I strongly believe bitcoin (as it currently stands) is NOT worth supporting, I recognize that people of good conscience can arrive at a different conclusion when it comes to supporting GPUGRID through D@H. But the current BOINC credit awards for nVidia participation in D@H are a perverse incentive to engage in a woefully inefficient activity. That Kepler may or may not remedy the problem in the future is a very poor argument for allowing the current situation to continue. | |
| ID: 23506 | Rating: 0 | rate:
/ Reply
Quote
| |
|
Toni Volunteer moderator Project administrator Project developer Project tester Project scientist Send message Joined: 9 Dec 08 Posts: 1006 Credit: 5,068,599 RAC: 0 Level Scientific publications | |
|
D@H is a different project, just like hundreds of others. Each has its credits/work ratio, for various reasons, mostly related to the impossibility to precisely measure flops performed. Credits in-equality between projects is not a new discovery. This may not be the right thread to discuss it. | |
| ID: 23516 | Rating: 0 | rate:
/ Reply
Quote
| |
|
GDF Volunteer moderator Project administrator Project developer Project tester Volunteer developer Volunteer tester Project scientist Send message Joined: 14 Mar 07 Posts: 1957 Credit: 629,356 RAC: 0 Level  Scientific publications | |
|
The credit at donateathome.org will be close to the one at gpugrid for the short application, even if the ops/s is higher over there because the algorithm is simpler and the gpu can perform better. A first adjustment is already inside the new beta D@H app which is going to be released soon there. | |
| ID: 23520 | Rating: 0 | rate:
/ Reply
Quote
| |
|
SMTB1963 Send message Joined: 27 Jun 10 Posts: 38 Credit: 524,420,921 RAC: 0 Level Scientific publications | |
D@H is a different project, just like hundreds of others. Each has its credits/work ratio, for various reasons, mostly related to the impossibility to precisely measure flops performed. Credits in-equality between projects is not a new discovery. This may not be the right thread to discuss it. While I tend to agree with your position on the value of credits, the reality is they do in fact serve as powerful incentives for how volunteers contribute their computing resources to GPUGRID and D@H. Arguments along the lines of "we can't address this issue because it's impossible to precisely measure flops performed" don't carry a lot of weight in my book. Why? GPUGRID currently uses credit incentives without any regard at all to flops/mips/whatever for long vs short and <24hr vs >24hr workunits. These bonuses have nothing to do with GPU performance and everything to do with affecting volunteer behavior. Presumably, the current GPUGRID bonus incentives exist to reward efficiency. If efficiency is a value that is shared by D@H, the possibility that there may be a problem with the way D@H credits are awarded relative to GPUGRID ought to be looked into instead of being dismissed out of hand. Further, there may be D@H volunteers running nVidia cards who might not realize that they are paying more money to their utility companies than they are generating for D@H. BOINC credits awarded at D@H don't seem to discourage this blatant inefficiency. | |
| ID: 23522 | Rating: 0 | rate:
/ Reply
Quote
| |
|
skgiven Volunteer moderator Volunteer tester Send message Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level Scientific publications | |
While I tend to agree with your position on the value of credits, the reality is they do in fact serve as powerful incentives for how volunteers contribute their computing resources to GPUGRID and D@H. For some people, and to varying extents, but not all. I've never been tempted by the maths, chess or similar projects; I wouldn't spend money on the electric to find a prime number, but if someone else want to, that's up to them. How many people do you think stopped crunching GPUGrid to crunch for D@H? Arguments along the lines of "we can't address this issue because it's impossible to precisely measure flops performed" don't carry a lot of weight in my book. Why? GPUGRID currently uses credit incentives without any regard at all to flops/mips/whatever for long vs short and <24hr vs >24hr workunits. These bonuses have nothing to do with GPU performance and everything to do with affecting volunteer behavior. If Boinc can affect the implementation of a universal GPU credit system that 'fairly' grants credit based on performance, and factors in considerations such as sp or dp, GPU utilization, CPU requirements, power requirements, other hardware requirements, bandwidth needs and so on, then great, they've got my backing 100%. The credit system here is based on steps (which reflects actual performance of the app on the GPU) and return time. Credit variation here exists due to the different task performances on the apps. GPUGrid usually runs several task types for different research projects. Presumably, the current GPUGRID bonus incentives exist to reward efficiency. Yes, and to pick up on concerns expressed earlier regarding the greenness of this system - running a high end card 24/7 is often twice as efficient (in terms of results per watt) as running lesser cards. Several of the crunchers here went well out of their way to produce performance per watt tables to encourage the purchase of greener cards for crunching. So yes, more credit is awarded for fast turnover of long tasks (which expedite the research over short tasks), and these tasks tend to be run on greener cards for crunching. If efficiency is a value that is shared by D@H, the possibility that there may be a problem with the way D@H credits are awarded relative to GPUGRID ought to be looked into instead of being dismissed out of hand. Further, there may be D@H volunteers running nVidia cards who might not realize that they are paying more money to their utility companies than they are generating for D@H. BOINC credits awarded at D@H don't seem to discourage this blatant inefficiency. The D@H credit system is more rudimentary. It rewards for task completion. The task runtimes are consistently similar on similar cards, so as with most projects faster cards get more credit, whether AMD or NVidia. Presently AMD cards are about three times as efficient generally. The problem with chaining the D@H credit system to reflect GPUGrid's is that they are two distinct projects with totally different tasks, apps and project needs. D@H has to work out what's right for D@H, and it will, in due course. Some credit changes are already planned there (comments and suggestions were considered, as normal), and the new credit system should go some way to alleviating your concerns. However there is a problem with making such changes now. The apps are very immature. Even within the first week the performance on the HD 7900's doubled. So credit for some card types doubled in a week. I don't want to see the credit change every few days. The 7770 and 7750's are AFAIK untested, and performances may change dramatically with drivers or app refinements, for these and other cards. Performances are discussed openly in the threads there. With so many cards coming onto the market over the next few months, the efficiencies (GPU running cost vs pooled BC credit) for different cards is largely unknown. Obviously NVidia cards are not efficient now, but the developers still need a benchmark to work from, or they won't be able to compare the new cards. I don't anticipate many NVidia's being used at D@H, at least until GF600 arrives. At present some of the AMD cards would at least on average match BC's to running cost, but only in some areas where electric costs are low. Then there are clock settings to consider; overclocking &/or under-volting the GPU/stream processors/GDDR changes the picture again. For sure the more capable crunchers can configure their card and settings to make them more viable. Everyone should remember it's an experimental project; everything is alpha. ____________ FAQ's HOW TO: - Opt out of Beta Tests - Ask for Help | |
| ID: 23528 | Rating: 0 | rate:
/ Reply
Quote
| |
|
SMTB1963 Send message Joined: 27 Jun 10 Posts: 38 Credit: 524,420,921 RAC: 0 Level Scientific publications | |
|
Thanks for your detailed and thoughtful reply, skgiven. It appears that you and I agree on a great many things, like:
Projects have discretion in determining the amount of credit awarded for task completion. BOINC credits may be granted based on factors other than GPU computing performance. BOINC credit influences volunteer behavior. Efficient utilization of volunteer computing resources is an important project objective. Project volunteers benefit from guidance on suitable hardware for a given project.
Do you believe D@H project leadership has a duty to inform volunteers of this possibility? Do you believe such a situation is inefficient? Do you believe that D@H should structure its BOINC credit incentives to discourage volunteers from deploying hardware unsuitable for bitcoin mining? ____________________________________________ | |
| ID: 23552 | Rating: 0 | rate:
/ Reply
Quote
| |
|
skgiven Volunteer moderator Volunteer tester Send message Joined: 23 Apr 09 Posts: 3968 Credit: 1,995,359,260 RAC: 0 Level Scientific publications | |
Do you believe that D@H volunteers could find themselves paying more in electricity costs than the value of the BTC they generate for the project? They could, and individually some inevitably will. What people contribute with is down to individual choice. At present its an early/experimental project, with a range of cards being used, but many of the top crunchers have high end cards which in theory could generate more funds than their running costs. Ideally, as a pooled group of contributors the total cost of electric would be less than the value of BitCoins. This would obviously go some way to making this approach of generating funds more acceptable to those that don't accept it. If 'in the long-run' the cost of generating revenue is less than the expense of fueling the method, it will have been very worthwhile. If not then it's just a novel way of generating funds. This is what make the experiment interesting and challenging. If you want to donate and have a top AMD card for crunching anyway, and use it for mining as part of a collective group that could well be a better way to make a donation - if collectively you pay less for the electric than the actual donation. I would not however exclude people from donating on grounds of their GPU. In many ways some of the older cards actual value is low, so it's really just down to cost of electric vs collective donation. For new cards that would exclusively be used at D@H then you would have to factor in purchase cards, but I doubt that many people will do that. Do you believe D@H project leadership has a duty to inform volunteers of this possibility? I think that has already been done. I would prefer that there was more detailed stats, but at an early alpha stage (with apps failing, development ongoing) any such stats or tables would be temporary, and it would probably take as long to generate such tables as it would to improve the code (which is the priority at this stage).
I don't know how efficient the cards are at this stage; that would depend on electric costs of every crunchers individual cards, as well as clocks, card specs and so on. We don't have all that info. Again, at this stage the cards will become more efficient with app development. So for example the doubling of performance of the 7900's may have made them efficient, but before that probably not. If performance improves again then their efficiency will improve too. New cards will change the whole picture. So telling someone to not crunch on a card now might be counterproductive in the long run if it impedes development.
The present situation is sufficient in my opinion; the more 'efficient' cards generate more credit anyway. I expect most high end NVidia crunchers will use their cards here or at other GPU projects. I have never had any inhibitions about telling someone their GPU specs are poor and 'not recommended' for the project, but at this stage I don't want to overly discourage the use of a card that might in the future be more efficient. While I don't think that present NVidia cards are 'efficient' or ever will be, it's down to an individuals choice and I'm slightly worried about making sweeping statements such as "NVidia cards give poor performance", not because it's inaccurate, but because that general impression might discourage people in a few months, possibly even for a GPU that is more efficient that AMD's (or not). Generic sweeping statements like that can pop-up in searches for years. I would encourage crunchers at D@H to make their own calculations of 'efficiency'. Lists can be maintained by the community. Some existing GPUGrid crunchers crunch at D@H, but mostly with AMD cards. Some former crunchers that presently just use AMD cards crunch at D@H, and some non-GPUGrid crunchers crunch at D@H (mostly AMD only crunchers). Overall there are not that many crunchers with credit, and there is not much work done on NVidia cards. ____________ FAQ's HOW TO: - Opt out of Beta Tests - Ask for Help | |
| ID: 23556 | Rating: 0 | rate:
/ Reply
Quote
| |
|
Mad Matt Send message Joined: 29 Aug 09 Posts: 28 Credit: 101,584,171 RAC: 0 Level  Scientific publications | |
|
I don't understand the logic here. This project never has been paying extensively high credits. Everyone has seen what happened at Primegrid last year, up to the point where folks started selling ATI cards and replaced them by Nvidias. | |
| ID: 23701 | Rating: 0 | rate:
/ Reply
Quote
| |
Message boards : Number crunching : Re Toni_TAKE3
