I'll come back later to comment about this. Just now I need to go and get some sleep. (There are a handful of us who are starting to fade badly in the push to get our new search up and running. It's a bit scary since I am spending hours logged in working as root on the main E@H server. Not very sensible...)
@ azwoody : Agreed, there are plenty of benchmarks about. The real trick is to get agreement about them ... which is why I hedged with 'sufficiently definite' .... :-)
@ Alinator: There is certainly a 'value' side to the coin which is independent of technicalities. Probably the Tar Pit analogy is even more apt! :-)
As it's pretty clear our dev's are taking a flogging at the moment with the workload, I'd wish to thank them for their efforts, whatever happens, and wish them a good sleep ( or three ... ).
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
After reading about how credits/cobblestones were adjusted for the new units I went back and did a basic comparison between what I was getting on the old units, the new units, and what I am getting at S@Home. I admit I'm not math genius so my results are really rudimentary but this is what I got as a result on the same system. I have rounded off here for simplicity sake.
E@home old unit - 9200sec 53credit divide credit by time = .005
E@home new unit - 47000sec 77credit result = .002
S@home - 17800sec 62credit result = .003
E@home old unit 2 1/2 hours 53credit
E@home new unit 13 hours 77credit
S@home unit 5+ hours 62credit
While I get units of varying times at Seti, I picked one that was as close as possible to an equivilant one in Einstein. I don't see where the new credits were brought into line with some of the other projects that Bruce mentioned in another thread. And if Seti was/is using Einstein's benchmarks for credit applied then it looks wrong to me based on my elementary calculations. Granted I don't run any other projects so I have no information on how they compare.
Maybe Bruce can explain what he meant by credit ratio being adjusted by 2.2 to bring it inline with other projects.
(edited to correct time of old unit hours under E@H)
Now I'm even more confused on the time/cobblestone comparison. Personally I'd prefer that all the projects give credit based on the same standard so that I'd know that no matter what science project I was working on would receive the same amount of credit based on the work.
One problem I see here comparing cr/time for Einstein with Seti (main site), is that the Seti application is open source. So on Seti, one can obtain an optimised version.
So do you compare Einstein with the Seti normal app, the optimised app or with a combination of both.
It is also not really fair to try and compare only on one type of cpu. As Einstein, on past experience, tends to work better with increases in clock speed, whilst Seti works better with increases in L2 cache.
Examples with Seti using optised apps (at same AR = 0.54) and SetBeta using standard app (credits increased to bring in line with Einstein)
[pre]Computer Einstein (54cr) Seti (52cr) SetiBeta (60cr)
AMD 2000+ (1.66GHz) 13k 14.6k N/A
Pent M 2.13GHz 11k 6.4k 13k
C2D 2.7Ghz 6.4k 4.5k 7.2k[/pre]
One can see that on Einstein the AMD, nearly 4yrs older than the Pent M, is nearly its equal, but on Seti it needs over twice the time to get approx same credits.
So please don't try to adjust on the results from one type of computer.
And for Bruce and his overworked developers please compare apples with apples, i.e. standard app to standard app.
With the new units here, it kind of backs up my "cpu time" is a good measure idea.
SR1 units on on specific box I have took 3h 10m (~52 credits)
It's now crunching some SR2, and while they've not completed, they'll be close to 15h, but I doubt they'll be worth ~250 credits!..
But, in cpu time, the measure would be the same for 5x sr1 vs 1x sr2 - 15 hours...
Do I do more "work" if I crunch for 15 hours on seti? Well, I still did 15 hours, so my "pay" should be the same. 2000 hours on CPDN vs 2000 hours on Seti? Same pay....
100 hours on a 500mhz box vs 100 hours on a 2.3 ghz box? Both worked the same amount of time.
Seti Classic has a great scheme (if you don't count cheaters!) where Wu's were counted, but that's not practial with Boinc... You got WU's that complete in minutes from one project, while other projects have WU's that take months...
You do a WU, it validates and you get 1 point added to your total for that project. It takes x hours to do that WU and you get x hours added to your total for that project. What could be more simple and easy to understand? Why must everyone try to compare apples to oranges and come up with what, a mess? The current credit system is a waste of time, always has been a waste, and always will be! It's all fluff and feel good numbers!
I have no idea how many WU's I do a day or what the average processing time is. How could I, all there is to work with is some silly number that changes as the credit goes up and down for no reason other than someone thinks it's out of line with what other projects are giving out! Forget the speedometer, give me the road miles and a clock!
"The guy that spends 60 minutes crunching on his 500mhz Pent on WU's gets the same "credit" as 60 minutes on a Pent 4 at 2gig crunching WU's..
They both contributed 1h of time... While others won't agree, this is a valid measure that can be used for any project on any HW/SW, and will level the playing field..."
Can't agree with that. When we add crediting system in projects we also start some sort of competition ... that competition force some people to have better result and more credits, so they start to invest to their computers even buying new one ... and spending more time involved in project, creating teams and animating people to join them ... in big part because competition.
All this increased power and number of computers and help doing research faster.
What going to happened if we stop count on that?
"The guy that spends 60 minutes crunching on his 500mhz Pent on WU's gets the same "credit" as 60 minutes on a Pent 4 at 2gig crunching WU's..
They both contributed 1h of time... While others won't agree, this is a valid measure that can be used for any project on any HW/SW, and will level the playing field..."
Can't agree with that. When we add crediting system in projects we also start some sort of competition ... that competition force some people to have better result and more credits, so they start to invest to their computers even buying new one ... and spending more time involved in project, creating teams and animating people to join them ... in big part because competition.
All this increased power and number of computers and help doing research faster.
What going to happened if we stop count on that?
Precisely. The people who don't care about credits will be quite happy with a count of hours, but by definition these people don't need to be considered in credit calculation because they don't care about credits. (Actually, WCG keeps totals of hours donated and WUs returned separately. It's not a bad idea.)
The people who care about credits care about them because they like to tinker with their machines to squeeze the maximum amount of work per hour of CPU time. To get these people to contribute we need a system of credits which is reasonably fair across projects. (At least fair enough so people won't crunch projects they are less interested in/supportive of just to get more credits.) It's a tough ask, but IMHO worth the effort to get as close as possible to the ideal.
Actually, WCG keeps totals of hours donated and WUs returned separately. It's not a bad idea.
That would be great. Then you could have whatever credit scam they come up with at the time, and, an actual count of WU's done along with the time spent doing them. Unless all the WU's are the same, and they never are, the only way to see how much difference a specific change to a machine makes is to keep track of the change in the number of WU's done per day over a period of time. That means you need a WU count and a tally of the time spent to do those WU's.
Something else that would be of value is an option to set cpu affinity. I've done enough WU's on SMP machine to know that it does make a difference. While it may not be that big a difference per WU, it adds up over time to make a big difference over all. If you do a lot of WU's over a give period of time, even a few minutes saved per WU adds up fast.
They've been trying to come up with "fair credits" since the birth of Boinc.. Cobblestone was based on math operations, then that went away, the credits, with correction factors..
Across a single project (with or without benchmark data) credits still don't seem to be a reflection of actual work done - but at best a guess.. Look at this project. For a given WU, all who crunched are given the same credit. As the project changes, the project uses a "correction factor" to award the estimated credit (that's why there's such a difference between SR1 WUs and SR2 WU's right now - it's been said "5 times the CPU time, and only 3 times the credit?"
Add into the mix, the numerous projects, each using their own "correction factor" for credits, and right now, today, credits can't even be used to "optomize" a box for boinc.. It's not apples and oranges, it's "apples, oranges, grapefruits, bananna, peaches, and grapes!"
A true measure, that is common between boxes in a given project, and common cross projects, is "total CPU time".
Any other scheme that's been tried in various projects over the last 3+ years, is only a "guess" at the work doin done....
And as far as contribution, isn't 1 hr of electricty on a 500mhz box like 1 hr of electricty on a 1ghz box?
RE: I'll come back later to
)
LOL... No problemo, been there done that!
Burnout is not a pretty thing. :-)
Alinator
@ Matt: LOL!! Clone on ...
)
@ Matt: LOL!! Clone on ... :-)
@ azwoody : Agreed, there are plenty of benchmarks about. The real trick is to get agreement about them ... which is why I hedged with 'sufficiently definite' .... :-)
@ Alinator: There is certainly a 'value' side to the coin which is independent of technicalities. Probably the Tar Pit analogy is even more apt! :-)
As it's pretty clear our dev's are taking a flogging at the moment with the workload, I'd wish to thank them for their efforts, whatever happens, and wish them a good sleep ( or three ... ).
Cheers, Mike.
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
After reading about how
)
After reading about how credits/cobblestones were adjusted for the new units I went back and did a basic comparison between what I was getting on the old units, the new units, and what I am getting at S@Home. I admit I'm not math genius so my results are really rudimentary but this is what I got as a result on the same system. I have rounded off here for simplicity sake.
E@home old unit - 9200sec 53credit divide credit by time = .005
E@home new unit - 47000sec 77credit result = .002
S@home - 17800sec 62credit result = .003
E@home old unit 2 1/2 hours 53credit
E@home new unit 13 hours 77credit
S@home unit 5+ hours 62credit
While I get units of varying times at Seti, I picked one that was as close as possible to an equivilant one in Einstein. I don't see where the new credits were brought into line with some of the other projects that Bruce mentioned in another thread. And if Seti was/is using Einstein's benchmarks for credit applied then it looks wrong to me based on my elementary calculations. Granted I don't run any other projects so I have no information on how they compare.
Maybe Bruce can explain what he meant by credit ratio being adjusted by 2.2 to bring it inline with other projects.
(edited to correct time of old unit hours under E@H)
Now I'm even more confused on the time/cobblestone comparison. Personally I'd prefer that all the projects give credit based on the same standard so that I'd know that no matter what science project I was working on would receive the same amount of credit based on the work.
One problem I see here
)
One problem I see here comparing cr/time for Einstein with Seti (main site), is that the Seti application is open source. So on Seti, one can obtain an optimised version.
So do you compare Einstein with the Seti normal app, the optimised app or with a combination of both.
It is also not really fair to try and compare only on one type of cpu. As Einstein, on past experience, tends to work better with increases in clock speed, whilst Seti works better with increases in L2 cache.
Examples with Seti using optised apps (at same AR = 0.54) and SetBeta using standard app (credits increased to bring in line with Einstein)
[pre]Computer Einstein (54cr) Seti (52cr) SetiBeta (60cr)
AMD 2000+ (1.66GHz) 13k 14.6k N/A
Pent M 2.13GHz 11k 6.4k 13k
C2D 2.7Ghz 6.4k 4.5k 7.2k[/pre]
One can see that on Einstein the AMD, nearly 4yrs older than the Pent M, is nearly its equal, but on Seti it needs over twice the time to get approx same credits.
So please don't try to adjust on the results from one type of computer.
And for Bruce and his overworked developers please compare apples with apples, i.e. standard app to standard app.
Andy
With the new units here, it
)
With the new units here, it kind of backs up my "cpu time" is a good measure idea.
SR1 units on on specific box I have took 3h 10m (~52 credits)
It's now crunching some SR2, and while they've not completed, they'll be close to 15h, but I doubt they'll be worth ~250 credits!..
But, in cpu time, the measure would be the same for 5x sr1 vs 1x sr2 - 15 hours...
Do I do more "work" if I crunch for 15 hours on seti? Well, I still did 15 hours, so my "pay" should be the same. 2000 hours on CPDN vs 2000 hours on Seti? Same pay....
100 hours on a 500mhz box vs 100 hours on a 2.3 ghz box? Both worked the same amount of time.
Seti Classic has a great scheme (if you don't count cheaters!) where Wu's were counted, but that's not practial with Boinc... You got WU's that complete in minutes from one project, while other projects have WU's that take months...
You do a WU, it validates and
)
You do a WU, it validates and you get 1 point added to your total for that project. It takes x hours to do that WU and you get x hours added to your total for that project. What could be more simple and easy to understand? Why must everyone try to compare apples to oranges and come up with what, a mess? The current credit system is a waste of time, always has been a waste, and always will be! It's all fluff and feel good numbers!
I have no idea how many WU's I do a day or what the average processing time is. How could I, all there is to work with is some silly number that changes as the credit goes up and down for no reason other than someone thinks it's out of line with what other projects are giving out! Forget the speedometer, give me the road miles and a clock!
"The guy that spends 60
)
"The guy that spends 60 minutes crunching on his 500mhz Pent on WU's gets the same "credit" as 60 minutes on a Pent 4 at 2gig crunching WU's..
They both contributed 1h of time... While others won't agree, this is a valid measure that can be used for any project on any HW/SW, and will level the playing field..."
Can't agree with that. When we add crediting system in projects we also start some sort of competition ... that competition force some people to have better result and more credits, so they start to invest to their computers even buying new one ... and spending more time involved in project, creating teams and animating people to join them ... in big part because competition.
All this increased power and number of computers and help doing research faster.
What going to happened if we stop count on that?
RE: "The guy that spends 60
)
Precisely. The people who don't care about credits will be quite happy with a count of hours, but by definition these people don't need to be considered in credit calculation because they don't care about credits. (Actually, WCG keeps totals of hours donated and WUs returned separately. It's not a bad idea.)
The people who care about credits care about them because they like to tinker with their machines to squeeze the maximum amount of work per hour of CPU time. To get these people to contribute we need a system of credits which is reasonably fair across projects. (At least fair enough so people won't crunch projects they are less interested in/supportive of just to get more credits.) It's a tough ask, but IMHO worth the effort to get as close as possible to the ideal.
--miw
RE: Actually, WCG keeps
)
That would be great. Then you could have whatever credit scam they come up with at the time, and, an actual count of WU's done along with the time spent doing them. Unless all the WU's are the same, and they never are, the only way to see how much difference a specific change to a machine makes is to keep track of the change in the number of WU's done per day over a period of time. That means you need a WU count and a tally of the time spent to do those WU's.
Something else that would be of value is an option to set cpu affinity. I've done enough WU's on SMP machine to know that it does make a difference. While it may not be that big a difference per WU, it adds up over time to make a big difference over all. If you do a lot of WU's over a give period of time, even a few minutes saved per WU adds up fast.
They've been trying to come
)
They've been trying to come up with "fair credits" since the birth of Boinc.. Cobblestone was based on math operations, then that went away, the credits, with correction factors..
Across a single project (with or without benchmark data) credits still don't seem to be a reflection of actual work done - but at best a guess.. Look at this project. For a given WU, all who crunched are given the same credit. As the project changes, the project uses a "correction factor" to award the estimated credit (that's why there's such a difference between SR1 WUs and SR2 WU's right now - it's been said "5 times the CPU time, and only 3 times the credit?"
Add into the mix, the numerous projects, each using their own "correction factor" for credits, and right now, today, credits can't even be used to "optomize" a box for boinc.. It's not apples and oranges, it's "apples, oranges, grapefruits, bananna, peaches, and grapes!"
A true measure, that is common between boxes in a given project, and common cross projects, is "total CPU time".
Any other scheme that's been tried in various projects over the last 3+ years, is only a "guess" at the work doin done....
And as far as contribution, isn't 1 hr of electricty on a 500mhz box like 1 hr of electricty on a 1ghz box?