Message boards : Number crunching : Fairer distribuiton of work(Flame Fest 2007)
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 . . . 10 · Next

AuthorMessage
River~~

Send message
Joined: 13 Jul 05
Posts: 456
Credit: 75,142
RAC: 0
Message 15088 - Posted: 13 Oct 2006, 5:46:21 UTC - in response to Message 15087.  
Last modified: 13 Oct 2006, 5:47:21 UTC

... what is the metric that shows one is "fairer" than the other? ...


Good points, but at present with only one suggested fix on the table, the debate would be more of philosophical interest.

For example if things were differemnt, if we had the luxury of choice between my quota limiting idea and the frequent suggestion to limit deadlines, quota limits would be favoured on the basis of fairness among machines regardless of speed; but dealine limiting would be favoured by those who felt that fairness lay in aiming to give each box an equal time slice from each release of work.

But sadly, both those options seem ruled out for technical reasons (see earlier posts in this thread).

That leaves only John's suggested tweak to the config file on the table. Whether you favour equality amongst boxes or equality on the basis of processing time, either way John's suggestion is better than the current situation.

I don't have a problem with having a debate between types of fairness (and if we do have that debate I declare an interest as owner of slow boxes) but please don't let that stop us asking the incoming admins to make the changes John suggests.

River~~
ID: 15088 · Report as offensive     Reply Quote
Gaspode the UnDressed

Send message
Joined: 1 Sep 04
Posts: 506
Credit: 118,619
RAC: 0
Message 15089 - Posted: 13 Oct 2006, 7:25:09 UTC - in response to Message 14951.  

Let's try again:

Since there is some debate about the use of deadlines for optimising the work flow see this excerpt from my earlier post:


It is CERN's problem if the the work takes too long to complete, and they are the people to decide if that is so, and to take action if it is needed. So, some time ago Chrulle implemented a dynamic deadline system which aims to optimise the return of work.



So, just to be absolutely clear on the point: Chrulle has implemented a deadline-based system to optimise the workflow.


Gaspode the UnDressed
http://www.littlevale.co.uk
ID: 15089 · Report as offensive     Reply Quote
River~~

Send message
Joined: 13 Jul 05
Posts: 456
Credit: 75,142
RAC: 0
Message 15091 - Posted: 13 Oct 2006, 13:41:05 UTC - in response to Message 15089.  

Let's try again:

Since there is some debate about the use of deadlines for optimising the work flow see this excerpt from my earlier post:


It is CERN's problem if the the work takes too long to complete, and they are the people to decide if that is so, and to take action if it is needed. So, some time ago Chrulle implemented a dynamic deadline system which aims to optimise the return of work.



So, just to be absolutely clear on the point: Chrulle has implemented a deadline-based system to optimise the workflow.



Yes indeed. Let's be clear.

Chrull has provided a deadline based system that worked well in the environment for which it was written, when a typical release of new work would last about a week on the server (ie there would be >1000 WU shown as available for a whole week.

If that environment returns, then the Chrulle solution is undoubtedly the best. Being from Chrulle it is both technically excellent and works in a reasonably fair way as between different hosts. I am not arguing otherwise, and I don't think anyone else in this thread has argued otherwise.

However, in an environnment where the entire release of work goes from the server in under an hour, our experience is that the Chruile solution fails to get the work back to the project as fast as possible and also fals to keep the donor community satisfied about its fairness. Both of those are observed facts. It is not Chrulle's fault that the environment is different now than when he produced his solution. Had he still been here, I feel sure he would have produced a new response to the new situation.

In the short term this is not a project issue, as the work is reutrned sufficiently soon for their purposes. In the long term it is a project issue, or at least in my view should be a project concern, due to the knock on effects on the morale of the participants. This opinion reflects my view that the voluntary work aspects of any DC programme are as important as the computer science and application science aspects.

So we clearly differ, but I am not sure at what point you differ from me Gaspode.

Are you saying the project should not react to a change in the circmstances?

Are you saying that in your opinion the voluntary work aspects are irrelevant to the long term success of a DC project? (if so I'd enjoy having that debate with you, but I'd prefer you to start another thread for it).

Or am I still missing your point?

best regards,
River~~
ID: 15091 · Report as offensive     Reply Quote
Philip Martin Kryder

Send message
Joined: 21 May 06
Posts: 73
Credit: 8,710
RAC: 0
Message 15093 - Posted: 14 Oct 2006, 0:52:34 UTC - in response to Message 15091.  

....fails to get the work back to the project as fast as possible and also fals to keep the donor community satisfied about its fairness....
River~~


Which of these two goals -
1) Work back as fast as possible.
or
2) Donor Community satisfied.

Do you feel more nearly defines fairness for you.
If it is the later, then I beleive my questions are not "merely" philosophical, but are the essence of the needed discussion. Any peceived satisfaction (or lack of it) should be based (in my opinion) on a clear understanding and concensus of what is meant by fairness.

If it is the former, (returning work as fast as possible), then I just plain disagree that speed of return is a measure of fairness.
It may well be that folks with slow machines should be "shut out" entirely - particularly since they have failed "to invest" in faster technology and should not be rewarded for having slower machine.

The deadline driven method, returns work "at the speed needed by the project" - which may be totally unfair, but which balances "fairness" with the scientific needs of the project...

ID: 15093 · Report as offensive     Reply Quote
Profile Keck_Komputers

Send message
Joined: 1 Sep 04
Posts: 275
Credit: 2,652,452
RAC: 0
Message 15097 - Posted: 14 Oct 2006, 11:06:09 UTC


1) Work back as fast as possible.
or
2) Donor Community satisfied.

Both of these goals are important and a balance needs to be achieved between them. Reducing the deadlines would be the most effective way to get the work back faster. It would cut out some of the slower hosts though.

Spreading the work out more would probably increase participant satisfaction, on the average. I would also expect it to reduce the time taken to get work completed. A task that is on a host being processed is more likely to returned sooner than one waiting in a queue no matter what the difference in speed between the hosts.
BOINC WIKI

BOINCing since 2002/12/8
ID: 15097 · Report as offensive     Reply Quote
River~~

Send message
Joined: 13 Jul 05
Posts: 456
Credit: 75,142
RAC: 0
Message 15098 - Posted: 14 Oct 2006, 17:15:23 UTC - in response to Message 15093.  


Which of these two goals -
1) Work back as fast as possible.
or
2) Donor Community satisfied.

Do you feel more nearly defines fairness for you.


Neither.

Work back as fast as possible could be a science goal (it is on some projects) but is not a participant goal. On this project it is not even a critical science goal, since the work gets back faster than the scientists/engineers are ready to use it anyway.

Donor community satisfied is an outcome of the majority of the donor community feeling that the way things are done is "fair enough". It is not a definition of fairness, as it could be achieved to a great extent using a number of different metrics. The aim would be to aim for an even handed distribution of work when work is scarce.

Whether "even handed" means equal work per box, per participant, or per gigiflop is a point for discussion, but the current situation does not achieve any of those. Personally if it went for any of those it would be "good enough" for me, and I believe for most crunchers.

So it is not a case of choosing one of your options over the other. It is a case for applying the "good enough" test to both sets of needs at once. On this principle, the focus is on improving whichever of the needs is currently met less well.

And the feeling of many users seems to be that the scrum for work does not meet the fairness test under any metric.

In contrast, an approach that prioritises one need over the other runs the risk of totally ignoring a possible huge improvement in the de-prioritised need for a tiny gain in the prioritised one.

That is the theory, in practice it is even more simple than that. John Keck points out we can improve both aspects at once. I don't understand how anyone can feel that is wrong?

R~~
ID: 15098 · Report as offensive     Reply Quote
KWSN - A Shrubbery
Avatar

Send message
Joined: 3 Jan 06
Posts: 14
Credit: 32,201
RAC: 0
Message 15099 - Posted: 14 Oct 2006, 21:00:50 UTC

River,

I have to say that your even approach and thoughful arguments are swaying my opinion.

NanoHive is currently working on a system to spread workunits among more hosts. Personally, I don't beleive that BOINC was designed to accomplish this task and get results back quickly, but we'll see what their project engineers can pull off.

If there is a way to do it server side (and from what I've read on various message boards it's not) their intention is a dynamic max results per CPU. Send out a limited number and not send more to that host until the previous results have been returned with no ultimate maximum per day.

In any case, it should be an interesting study.

ID: 15099 · Report as offensive     Reply Quote
Profile anarchic teapot
Avatar

Send message
Joined: 15 Feb 06
Posts: 67
Credit: 460,896
RAC: 1
Message 15100 - Posted: 14 Oct 2006, 21:15:39 UTC - in response to Message 15097.  


1) Work back as fast as possible.
or
2) Donor Community satisfied.

Both of these goals are important and a balance needs to be achieved between them. Reducing the deadlines would be the most effective way to get the work back faster. It would cut out some of the slower hosts though.


I have to agree with both Keck and River~~ on this one. Even if getting the work back within a shorter time period isn't currently a priority, it may be some day; in fact, probably will be by past performance since we've already had the situation where new WUs can't be sent out until the results from the previous batch are returned and there's been a systematic 10-day gap between the last WU going out and it being returned.

And of course, the important bit abut keeping the DC satisfied is that if it isn't, its component beings will leave the project: a serious handicap for future computing power.

I do like the idea of a maximum number of outstanding WUs per CPU, and I think many BOINC projects would find such a load-spreading system useful.
sQuonk
Plague of Mice
Intel Core i3-9100 CPU@3.60 GHz, but it's doing its bit just the same.
ID: 15100 · Report as offensive     Reply Quote
Philip Martin Kryder

Send message
Joined: 21 May 06
Posts: 73
Credit: 8,710
RAC: 0
Message 15101 - Posted: 14 Oct 2006, 21:17:41 UTC - in response to Message 15098.  
Last modified: 14 Oct 2006, 21:48:38 UTC


Which of these two goals -
1) Work back as fast as possible.
or
2) Donor Community satisfied.

Do you feel more nearly defines fairness for you.


Neither.
...
Donor community satisfied is an outcome of the majority of the donor community feeling that the way things are done is "fair enough". It is not a definition of fairness, as it could be achieved to a great extent using a number of different metrics.
...


I disagree.
There are many folks who actively measure things like customer satisfaction - in this case "donor" satisfaction.
Surveys could measure peceived fairness as a sub-component of "satisfaction".

.... The aim would be to aim for an even handed distribution of work when work is scarce.
.... It is a case for applying the "good enough" test to both sets of needs at once.
....

I believe your vision of a "good enough" test is in fact a form of survey (of you or whomever you ask) as to whether or not you are satisfied.



ID: 15101 · Report as offensive     Reply Quote
Philip Martin Kryder

Send message
Joined: 21 May 06
Posts: 73
Credit: 8,710
RAC: 0
Message 15102 - Posted: 14 Oct 2006, 21:54:24 UTC - in response to Message 15097.  


....
A task that is on a host being processed is more likely to returned sooner than one waiting in a queue no matter what the difference in speed between the hosts.


Well, all computers wait at the same speed - even the fastest.
That's for sure.

But a machine with 5 units of work can finish them faster than a machine one-tenth of its speed can finish even a single work unit.



ID: 15102 · Report as offensive     Reply Quote
Profile Maverick

Send message
Joined: 10 May 06
Posts: 8
Credit: 2,927
RAC: 0
Message 15103 - Posted: 15 Oct 2006, 15:52:02 UTC - in response to Message 15102.  

[quote]


But a machine with 5 units of work can finish them faster than a machine one-tenth of its speed can finish even a single work unit.



This maybe right, but even the slowest machine would have done some wu's since the last were given out. There are some people grabbing all the work and others don't get any.
Sincerly
Maverick
Caliban
ID: 15103 · Report as offensive     Reply Quote
River~~

Send message
Joined: 13 Jul 05
Posts: 456
Credit: 75,142
RAC: 0
Message 15105 - Posted: 15 Oct 2006, 17:55:21 UTC - in response to Message 15101.  
Last modified: 15 Oct 2006, 17:55:48 UTC


There are many folks who actively measure things like customer satisfaction - in this case "donor" satisfaction.
Surveys could measure peceived fairness as a sub-component of "satisfaction".


That is different from how I understood you at first, and I would still say that satisfaction does not define fairness. You original phrasing suggested to me that fairness was the only component of donor satisfaction, or maybe the only significant one. I don't accept that, but maybe it was not what you meant anyway.

Your new choice of words I agree with. Perceived fairness is a sub-component of satisfaction. More to the point, perceived unfairness leads directly to dissatisfaction -- not just in DC projects but amongst humans generally. If it feels fair, it is hardly noticed; if it feels unfair then does everyone get to know!


I believe your vision of a "good enough" test is in fact a form of survey (of you or whomever you ask) as to whether or not you are satisfied.


It is certainly something that can be measured formally by a survey. I'd say that you can get a feel for it without anything as formal as a survey -- from comments on a messaging board for example. I would not call a brief scan of these boards a survey, but perhaps we are just using the word differently.


Thanks to all contribuotrs who have pulled me up on details - especially those who have disagreed with me! I understand my own opinon more clearly as a result.

River~~
ID: 15105 · Report as offensive     Reply Quote
Philip Martin Kryder

Send message
Joined: 21 May 06
Posts: 73
Credit: 8,710
RAC: 0
Message 15107 - Posted: 15 Oct 2006, 19:51:07 UTC - in response to Message 15105.  
Last modified: 15 Oct 2006, 20:16:43 UTC


There are many folks who actively measure things like customer satisfaction - in this case "donor" satisfaction.
Surveys could measure peceived fairness as a sub-component of "satisfaction".


That is different from how I understood you at first, and I would still say that satisfaction does not define fairness. You original phrasing suggested to me that fairness was the only component of donor satisfaction, or maybe the only significant one. I don't accept that, but maybe it was not what you meant anyway.

Your new choice of words I agree with. Perceived fairness is a sub-component of satisfaction. More to the point, perceived unfairness leads directly to dissatisfaction -- not just in DC projects but amongst humans generally. If it feels fair, it is hardly noticed; if it feels unfair then does everyone get to know!






I believe your vision of a "good enough" test is in fact a form of survey (of you or whomever you ask) as to whether or not you are satisfied.


It is certainly something that can be measured formally by a survey. I'd say that you can get a feel for it without anything as formal as a survey -- from comments on a messaging board for example. I would not call a brief scan of these boards a survey, but perhaps we are just using the word differently.


Thanks to all contribuotrs who have pulled me up on details - especially those who have disagreed with me! I understand my own opinon more clearly as a result.

River~~



Thanks -
Nice reviewing the issue with you - and helping me understand better.

Over time, I would expect that folk would join or leave a project based on "satisfaction".

One danger of "scanning the message boards" is that the posters "self select" and may not represent a useful cross section.

I didn't choose to leave LHC because of lack of work.
Rather, I boost it to 99.9% resource share and run other stuff (Rosetta).
I figure as long as the work gets done as needed by the project it doesn't matter who does it. -Later when there in more work, I'll get some.

I have turned off another project due to "dissatisfaction" because the WorkUnits were very, very long and non-deterministic in lenght.



ID: 15107 · Report as offensive     Reply Quote
Philip Martin Kryder

Send message
Joined: 21 May 06
Posts: 73
Credit: 8,710
RAC: 0
Message 15108 - Posted: 15 Oct 2006, 20:22:48 UTC - in response to Message 15103.  

[quote]


But a machine with 5 units of work can finish them faster than a machine one-tenth of its speed can finish even a single work unit.



This maybe right, but even the slowest machine would have done some wu's since the last were given out. There are some people grabbing all the work and others don't get any.
Sincerly
Maverick


I think that this points to the concept of "fairness" being:
"give at least some work to as many machines as possible."

The good side of this is that it will spread interest in the project as widely as possible.

ID: 15108 · Report as offensive     Reply Quote
Gaspode the UnDressed

Send message
Joined: 1 Sep 04
Posts: 506
Credit: 118,619
RAC: 0
Message 15110 - Posted: 16 Oct 2006, 17:34:53 UTC - in response to Message 15091.  
Last modified: 16 Oct 2006, 17:40:52 UTC



Yes indeed. Let's be clear.

Chrull has provided a deadline based system...

So we clearly differ, but I am not sure at what point you differ from me Gaspode.

Are you saying the project should not react to a change in the circmstances?

Are you saying that in your opinion the voluntary work aspects are irrelevant to the long term success of a DC project? (if so I'd enjoy having that debate with you, but I'd prefer you to start another thread for it).

Or am I still missing your point?

best regards,
River~~


I am not sure we do differ. I have had the debate about participant support in the SETI forums when BOINC was still V3, and raised the issue with David Anderson because support was diabolically bad at that time. Poor support from CERN is one reason why I have withdrawn processing power from the project for the time being. It is true that the available LHC workload doesn't distribute evenly across the available participant base and on that I have already made my position clear. Since there is presently more crunching power than available work is my second reason.

(Incidentally,I subsequently formed the view that the approach adopted by SETI@Home is scientifically pointless, and took my processors elsewhere. This is not my view of LHC@Home)

As to whether LHC@Home should adapt its modus operandi, then it should if it sees the need. Of course, it presupposes that they are able to change, and with no staff assigned nothing at LHC will change until the migration to QMC is complete. If you read the posts in this thread you will see that a number of contributors are suggesting that control of deadlines might be away to 'improve' matters. I am merely highlighting the fact that this has already been done.

The same posts also take 'improve' in the sense that benefits the participant base as they see it. This doesn't necessarily tally with what the project would see as an improvement, and since I am here for the science (when I AM here) it is surely the project that should decide what constitutes an 'improvement'.


Gaspode the UnDressed
http://www.littlevale.co.uk
ID: 15110 · Report as offensive     Reply Quote
Dotsch
Avatar

Send message
Joined: 7 Aug 05
Posts: 60
Credit: 73,746
RAC: 0
Message 15112 - Posted: 16 Oct 2006, 20:03:15 UTC - in response to Message 14917.  

I think, it is a good idea to reduce the max. cache time. It will be fairer for all users.

ID: 15112 · Report as offensive     Reply Quote
Philip Martin Kryder

Send message
Joined: 21 May 06
Posts: 73
Credit: 8,710
RAC: 0
Message 15114 - Posted: 17 Oct 2006, 0:44:55 UTC - in response to Message 15110.  
Last modified: 17 Oct 2006, 0:46:47 UTC

[quote]
..., and since I am here for the science (when I AM here) it is surely the project that should decide what constitutes an 'improvement'.



I always took that view as well -

However, I'm coming around to the "if folk get discouraged and leave, that isn't good for the project (or the science) either," point of view.

I think that as long as there is a surfeit of folk willing to contribute, and a shortage of work, this is all moot from a science viewpoint.

However, At some point, it may happen that there is work to do, but all the volunteers have left....


ID: 15114 · Report as offensive     Reply Quote
Profile Maverick

Send message
Joined: 10 May 06
Posts: 8
Credit: 2,927
RAC: 0
Message 15115 - Posted: 17 Oct 2006, 7:58:21 UTC - in response to Message 15114.  

[quote][quote]


However, At some point, it may happen that there is work to do, but all the volunteers have left....


I'm a little sad, because I haven't got any work for months and now I see there is only work for some crunchers who seem to be better than others. If I don't get any wu's the next time new one's are to crunch, I'll forget about LHC.

Sincerly
Maverick
Caliban
ID: 15115 · Report as offensive     Reply Quote
River~~

Send message
Joined: 13 Jul 05
Posts: 456
Credit: 75,142
RAC: 0
Message 15116 - Posted: 17 Oct 2006, 18:52:05 UTC - in response to Message 15110.  

...
The same posts also take 'improve' in the sense that benefits the participant base as they see it. This doesn't necessarily tally with what the project would see as an improvement, and since I am here for the science (when I AM here) it is surely the project that should decide what constitutes an 'improvement'.


I only partly agree.

There is what is felt to be good by the users, one aspect. Only the users can say what that is.

There is what is needed by the scientists / engineers in terms of accuracy, turnround etc. That can only be specified by the scientists.

Both communities need to be able to say not only what their ideal is, but also where the "good enough" line comes.

Having heard all of that, then and only then is it appropriate for the project admins/owners to decide. That may not be the scientists, in the case of CERN the ultimate owners are the European taxpayers (not just from EU countries). In the case of US based projects it often is the scientist who owns the project, at least while the funding lasts (this is due to a fundamental difference in the way funding works acroos the pond, and further comment would be off topic here).

Ultimately it is the project owners' decision, but only after understanding the needs / wishes of both the donor and recipient communities

It is true we don't have the right to insist, but in my view if we donate work we do earn the right to be heard and have our collective views taken seriously even if the final decision goes against us.

Ultimately it is still up to the project, but that is a form of giving the project "the right to decide", but in a very different context than when that same phrase is used to try to stifle the expression of participant discontent. (I do not say you use it with that intent)

R~~
ID: 15116 · Report as offensive     Reply Quote
Profile FalconFly
Avatar

Send message
Joined: 2 Sep 04
Posts: 121
Credit: 592,214
RAC: 0
Message 15118 - Posted: 17 Oct 2006, 22:39:27 UTC - in response to Message 15116.  
Last modified: 17 Oct 2006, 22:42:32 UTC

I'm periodically checking with LHC to see how it is doing.

Looking at the last months, I'd say :
There is basically no "fair" distribution of work possible, given the tiny amount of Work that pops up once a while.

Even if every User was to be handed one exclusive WorkUnit, it would still leave others without work...

Therefor, the whole discussion got a bit obsolete IMHO, as the current Project needs stand versus a Community with over 100-fold the required computing power...

My bet is, once the change in infrastructure are complete and new goals are set (and some actual staff getting back to it), the whole 'problem' will simply vanish as fast as it arised.

Ironically (while I always liked LHC), other Projects - established and emerging - are literally dying for more computing power as we speak, plus people can witness a direct line to the staff that actually realizes Community inputs where possible. Just a point to consider.
Scientific Network : 45000 MHz - 77824 MB - 1970 GB
ID: 15118 · Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · 4 · 5 . . . 10 · Next

Message boards : Number crunching : Fairer distribuiton of work(Flame Fest 2007)


©2024 CERN