Why bother overclocking?

[Slicker]

I keep talking about K.I.S.S. and why I think BOINC is headed in the wrong direction. Here's one more reason. I was reading through the server_stable release notes this morning and came across this little note about a fix that I don't recall ever being discussed on the boinc_dev mailing list:


David 9 Jan 2012
- scheduler: the p_fpops value reported by clients can't be trusted.
Some credit cheats (e.g. with credit_by_runtime) can be done
by reporting a huge value.
Fix this by capping the value at 1.1 times the 95th percentile
of host.p_fpops, taken over active hosts.

No wonder the credit is so crappy on many projects. Anyone with a machine that doesn't fit the norm is assumed to be a cheater. What's the point of overclocking by 25% or more if you are going to be limited to 1.1 times the credit of that of the 95th percentile. Assuming that less than 5% of people crunching overclock their machines (e.g. the masses of set and forget certainly aren't overclocking!), then the 95th percentile is a stock machine. The assumption is that the same machine running on all operating systems returns the same GFLOPS calculation which it doesn't. Then again, isn't this a catch-22? If a machine is overclocked, it doesn't count towards the top 5% because it is assumed to be cheating and the number is capped at 1.1 x stock when it should instead be raising the bar and causing the value for the 95th percentile to be increased.

Note that it doesn't say the 95th percentile of MATCHING hosts either. So, if you have a new top of the line computer, you also get screwed because the majority of people will have much slower computers and your credit then gets capped because of their average.

Or, am I looking at this one all wrong?

[zombie67]

[...] and came across this little note about a fix that I don't recall ever being discussed on the boinc_dev mailing list

Of course not. Nothing is EVER discussed prior to implementation. DA's opinion is the only one that matters. Or at least it's the only opinion that DA will listen to. The only discussion that goes on, is dealing with problems after the fact.

It's been pretty fun lately, watching all the other admin/developer folks bitch slapping DA for moving from SVN to git, without telling anyone first. Even Eric Korpela (SETI@home admin) was throwing punches.

[Beerdrinker]

It seems to me, that they are digging their own grave? If people with high-end systems get slapped for having such a system, why bother participating in BOINC?

[kaptainkarl1]

When did communism become vogue?

[Duke of Buckingham]

http://1.bp.blogspot.com/_m_3PqTnwV8k/S6KjlnVKtlI/AAAAAAAAO10/catDyQabCO8/s720/Du+Juan+-+Vogue+China+April+2010+-+3.JPG
VOGUE and Communism

[trigggl]

I guess I better keep an eye on the credit reward differences between my two 3GHz computers. One of them is 1.8GHz stock.

[DrPop]

Oh man...this is insane. You mean I've been wasting probably hundreds of watts by O/Cing my rigs for how long??? Obviously this must not count toward GPUs, because I see a big increase in credit when I O/C mine. But CPUs...man, they suck down the juice when O/Ced, and I'll be awful sore if we aren't getting paid back appropriately for our work. What kind of a ship are they trying to run here, anyway? Aren't they wanting MORE folks to come on board the Distributed Computing club?:confused::mad:

[Fire$torm]

Oh man...this is insane. You mean I've been wasting probably hundreds of watts by O/Cing my rigs for how long??? Obviously this must not count toward GPUs, because I see a big increase in credit when I O/C mine. But CPUs...man, they suck down the juice when O/Ced, and I'll be awful sore if we aren't getting paid back appropriately for our work. What kind of a ship are they trying to run here, anyway? Aren't they wanting MORE folks to come on board the Distributed Computing club?:confused::mad:

Only projects that use CreditNew. Best just to avoid them unless your into M&Ms... :P

[Bryan]

The issue of OCing or not isn't quite as straight forward as it appears. If you have a project that has WU where the amount of computing varies greatly then it makes sense to pay X credits per flop times the amount of time it takes to finish a WU. The 95th percentile is probably approaching the high end machines. So if you are OCing 25% and they are paying 10% above the 95th percentile computers then you may be losing 15% per WU. Your net gain however will be higher than the 10% level because you will turn in MORE WU by OCing to 25% than if you dropped to a 10% OC. Ye olde Catch 22!

Where you could really get screwed would be on something like the old Aqua that had a multithreaded CPU app. If the majority of the folks were running quads or early V8s and you throw down a 6 core AMD or 6 core HT Intel then you might really get screwed. You'd still make more credits by OCing to the hilt but only becaues you turned in more WU.

This could also affect GPU projects like Moo where they use all GPUs. Just imagine if that was implemented on Moo and you put two 7990s (when they come out) in a single machine. Talk about blowing away the 95th percentile!!!!

As usual, DA has his head where the sun don't shine. This gets back to his philosophy (my interpretation) that 24 hours of computer/GPU time should pay the same regardless if you are running a 6 core or 1 core machine! In his eyes an abacus would qualify! Socialistic crunching ya know!

[kaptainkarl1]

Only projects that use CreditNew. Best just to avoid them unless your into M&Ms... :P

Which projects run the Credit Poo? I would very much like to avoid them if at all possible or at least until this DA character drops dead and reality sets in again.

[Slicker]

Which projects run the Credit Poo? I would very much like to avoid them if at all possible or at least until this DA character drops dead and reality sets in again.

I'm pretty sure all of the ones participating in the EURO Crunch 2012 Series use CreditNew. It is hard to say for sure because there isn't a version x.y.z for the server code, but rather changset numbers and the numbers change daily. While they date the changes, one can't tell from the changeset number when it was last changed. All I can say for sure is that even though DA made the code change, only projects which have upgraded their servers since that code change are now using that new logic.

DA must watch COPS where are the criminals are idiots. His "fix" assumes that the person cheating who has all the tools necessary to build his own version of the BOINC client, is dumb enough to only increase the flops count. Yep. Like the hacker woudn't know that a P4 with a bazillion gflops is not normal but a 16 processor multi-core Xeon server would be. All they have to do is change the open source code to pretend to be a better machine and who would know the difference? No one. Not unless the "95th percentile" is ALL active hosts and not active hosts with the same type and number of processors. If that's the case, then the guys with those boxes truly are getting shafted. That's what worries me about this change.

[kaptainkarl1]

I'm pretty sure all of the ones participating in the EURO Crunch 2012 Series use CreditNew. It is hard to say for sure because there isn't a version x.y.z for the server code, but rather changset numbers and the numbers change daily. While they date the changes, one can't tell from the changeset number when it was last changed. All I can say for sure is that even though DA made the code change, only projects which have upgraded their servers since that code change are now using that new logic.

DA must watch COPS where are the criminals are idiots. His "fix" assumes that the person cheating who has all the tools necessary to build his own version of the BOINC client, is dumb enough to only increase the flops count. Yep. Like the hacker woudn't know that a P4 with a bazillion gflops is not normal but a 16 processor multi-core Xeon server would be. All they have to do is change the open source code to pretend to be a better machine and who would know the difference? No one. Not unless the "95th percentile" is ALL active hosts and not active hosts with the same type and number of processors. If that's the case, then the guys with those boxes truly are getting shafted. That's what worries me about this change.

Don't you have to turn in work to get credits? How does one make a quad core process more work units correctly and get the credits for it?

[Rattledagger]

I'm pretty sure all of the ones participating in the EURO Crunch 2012 Series use CreditNew.
CPDN doesn't use CreditNew, they're relying on the trickles as they've always done.


As for CreditNew, don't know exactly how it's working, or for that matter supposed to work...

... But, atleast if it wasn't just an artefact of not being "trusted" then downloaded the work for the pentathlon in WCG, atleast to me it looked like C4CW gave the same credit per wu for all work downloaded at the similar time.

So, if this wasn't just a fluke, overclocking would obviously have an effect, since 25% faster would mean 25% more credit, not per wu, but per day.

[Mad Matt]

I think this is another brilliant example why DA does not have any interest in speeding up science. What can you do if folks run a science project who do not care about speeding up the process and using resources in the best possible way? Little. It's futile to solve this with any logical approach. The only theory springing too my mind is: it does not matter how long SETI will not find a thing... BOINC should be run by somebody from a real science project who has urgent interest in scientific results.

After all, this is bitter for the CPUs, but GPUs will be less affected. Folks develop GPU apps to speed up things. CreditNew did not last for more than a few weeks at POEM after introducing the GPU app. So at least the GPU projects will always find a workaround, in their own vital interest.

Not really a solution, but at least good to know.

[Slicker]

Don't you have to turn in work to get credits? How does one make a quad core process more work units correctly and get the credits for it?

Let's say you have an E2160 that runs at 1.8Ghz. After changing the BOINC client code to report your machine as much faster CPU (a.k.a. 3.6Ghz and double the GFLOPS) you turn in a WU that takes 1 hour. BOINC looks up how much credit to award a 3.6Ghz CPU for an hours worth of work and grants the credit. Because BOINC thinks you have a faster CPU than you really do, it will grant double the credit it should. Or, it may average the credit it thinks you should get with your wingman's and you get 50% more than you should. Either way, you get more credit for each WU returned than if BOINC thought you had a slower machine.

That can't happen on projects which grant X credits per WU because on those projects, you get the same credit for each WU regardless of how long it took to crunch or what your computer claimed. Optimized apps or overclocking on those projects means more credits per CPU second. But, some projects awarded higher credit than projects. So, DA introduced CreditNew so that credit wasn't handled by the project admins any more but rather via DA's algorithm which takes into account the credit from all the other projects for similar types of hosts. Unfortunately, that means that if the Collatz ATI app runs 96 times faster than the CPU app, it won't get 96 times the credit. Rather, since the average GPU app is only 10 times faster than the CPU apps, it will only get 10 times the credit even though it finishes 96 times faster. That's considered fair in CreditNew. Likewise, if the GPU app only utilizes 50% of the GPU and uses an entire CPU as well, CreditNew awards the same credit as projects which utilized 98% of the GPU and use less than 1% CPU. That's why I don't like CreditNew.

[kaptainkarl1]

If that's the case then I say we overclock DA with a couple of nice heavy grandfather clocks so he gets the point.

[Aux10]

Why bother participating in BOINC?

For the science! Or so most of the newbies say.

[Slicker]

For the science! Or so most of the newbies say.

The science angle also helps when the wife sees the monthly electric bill which is always over $300.

[Rattledagger]

For the science! Or so most of the newbies say.
Oh, some old-timers also cares much more for the science than the worthless (atleast to me) points in some boring math- or crypto-project. (:| But, I do like a challenge, so did still run Collatz during the pentathlon and nearly doubled my collatz-points.

[Aux10]

Oh, some old-timers also cares much more for the science than the worthless (atleast to me) points in some boring math- or crypto-project. (:| But, I do like a challenge, so did still run Collatz during the pentathlon and nearly doubled my collatz-points.

You just haven't been bitten yet! ;)

[rgathright]

Socialistic crunching ya know!
All I have for people who preach fairness is... get a load of this!
**==**==**==**==**==**==**==**==**==**==**==