Poaching

[Prime95]

Originally Posted by outlnder

If we are to place limits on how long an exponent may be checked out to any individual, I vote for 6 months MAX.

Consensus seems to be building around a sliding scale. It's 2 to 3 months for the smallest double-checks and first-time tests (to avoid holding up milestones), 6 months for recycled exponents, 12 months for an exponent at the leading edge. 2+ years for a 33M exponent.

Give or take.

[garo]

Yeah,
That sounds like a plan. It may also be useful to tune the expiration times for the leading edge of LL and DC separately. Looks like first time tests go for leading to trailing in a significantly longer time than doublechecks. Still I guess if you weren't able to complete in one year, you can't hold on to the exponent any longer.

[QuintLeo]

I can see the second checks being held to a shorter timeframe, since they hold up milestones more.

I just think it's a bad idea to say "your machine can't finish an exponent in a year, go away" when a LOT of relatively current machnes can't manage that.

Heck, it's only been a year or so since 500ish Mhz machines were still considered 'current' (my last girlfriend bought a brand new eMachines Celeron-533 from Best Buy at that time), abet at the entry level. 2 years ago, they *were* state-of-the-art, and 3 years ago they were LEADING edge. Not all that old, folks - most such machines are probably still in use, and a majority of them are probably *still* the only machine the owner has.

To put the above into perspective - I've got a few 450-550 Mhz K6 and Celeron boxen - the K6 boxen I've mostly got set to factoring, 'cause the FPU on those is pretty wimpy. The Celerons I mostly have set for first-time testing - and those are mostly looking at a little over a year per exponent.

I'm not suggesting at all that there should be *no* time limit - I just think this "3 month" or "6 month" nonsence is TOO short. The majority of folks can't *afford* to upgrade their machine every other year or more often.

First checks should be given a significantly longer time limit, and the limit *should* slide higher as the exponent gets bigger. Factoring shouldn't be given a limit at all, IMO - even my K5 boxen manage to factor an exponent in a month or two, and those *are* way old outdated boxes. Second Checks should have a shorter time limit since they're the most likely place that will impact verification of a milestone.


Your opinions probably differ.

9-)

[outlnder]

I'm sorry Quint, but I would not waste the electricity on a machine that took more than 30 days on a first time LL test.

I am an unemployed pharmer and my slowest machine is a 1.1 GHz Duron.

Sometimes you just hafta let the old machines die.

[Prime95]

Originally Posted by QuintLeo

I just think it's a bad idea to say "your machine can't finish an exponent in a year, go away"

I don't think we are saying that. A leading edge first time test today is unlikely to hold up a milestone for maybe 2 years. I'm not advocating yanking a reservation just because you've had it one year.

I think we are proposing reassignment if you take more than a year and some other criteria is met such as:
a) You aren't making significant progress.
b) You are holding up a milestone.
c) Require the user to fill out a web form saying "I'm still working on it"

I think adding smarts to the server will keep 99% of the people happy. We just need to program the server such that your Celeron gets a leading edge first time test.

[binarydigits]

I didn't bother checking, but I did assume there was some poaching going on when I saw how fast those trailing DCs were dropping off the world test status.

Quint: WRT "Factoring shouldn't be given a limit at all": You need to take a look at the three factoring jobs between 14.9M and 15.1M which have been running for about two years or more, two of which are not yet halfway finished! My recently deceased P90 laptop would have run them faster than that even if it were only on for a couple of hours per day.

outlndr: Maybe my electricity is cheaper than yours, but I still have a lot of slow machines running GIMPS, and they will do so until they die. (Also, my home central heating unit is broken, and without all the PCs running I would probably freeze to death. ;) ) Of course, I do make sure that they are all running appropriate work. Last year I adjusted some ~400MHz boxes from LL to DC, and a couple even slower from DC to TF.

It is unfortunate that the work assignment types are controlled by the client rather than the server - but I am sure that can be fixed with new server software that will override the client.

I have a feeling that most of the 'trailing edge' jobs have been forgotten by their owners. I agree with George that time constraints should be used only in specific areas, and not just in general. There are instances where it is fairly obvious that it is needed. Some sort of assignment management will surely cut out most of the poaching, if not all.

[QuintLeo]

Outliner - a Celeron or K6 500Mhz machine doesn't use a lot of electricity, probably half or possibly LESS that used by your Duron 1.1, and usually a LOT less than an Athlon or P4.

For comparison - my #1 machine is an Athlon XP1800+ running at about 1600 Mhz (overclocked). It eats right about 150 watts. One of my *dual* Celeron BP6 machines runs a pair of Celeries at 600 Mhz each, and eats a little LESS than 100 watts. The Athlon box does a LL on a 33M exponent at .258 per iteration, each CPU in the BP6 does a 33M LL at about 1.7ish per iteration - despite the memory overhead of both CPUs doing LLs. I suspect if I swapped one back to trial factoring, the other would drop the per-iteration down to around 0.9ish - which is pretty close "per watt" to my Athlon. Yet each CPU in the BP6 box is going to take well over a year to LL one exponent, while the Athlon looks to average one every 4 months or a little more (depends on how much time I spend doing other stuff on that machine).

By *your* stated standard, your Duron isn't efficient and is a waste of electricity compared to a P-IV - uses almost as much electric for a LOT slower crunch rate (unless it's a slow P-IV doing trial factoring).

Most distributed projects aren't about "efficiency" - they're about farming out the work as widely as possible to get the work done as fast as possible. If that means a few small *sections* of the work don't go all that fast, so what? It's not like a Mersenne Prime is going to go away just 'cause it doesn't get discovered the first month it's tested.

Just because *you* don't think it's worth the electric does NOT mean that's the case for other participants. Especially for those of us in *cold* climates, or places that have cold winters - right now, the electric usage of my machines IS NOT A FACTOR - none of it is going to waste, as any electric turned into heat by my computers is that much *less* heat that has to be generated some other way. For me, and for a lot of other folks, it *is* efficient to run older slow machines - the electric would be TOTALLY wasted otherwise.

[outlnder]

Quint, you are absolutely correct. I apologize.

I quess the difference is that in California the cost of electricity is 2 to 5 times what you pay. We have tiers that increase exponentially as usage increases.

Most families with four people are being penalized multiple times for just using the normal amount of electricity for those 4 people. You can imagine how big the penalty is for a single peson in a one bedroom apartment using twice the electricity of that family of four.

[QuintLeo]

Hey, viewpoints vary, as do electric costs. I'm on "time of day" rate where I'm at, which means 19 hours out of the day my electric is VERY cheap, and the other 5 it's probably about 60-70% of the over-expensive California cost. (I do remember San Diego electric pricing, spent almost 12 years out there - it was a PLESENT shock to get my first Indianapolis electric bill, and the power where I'm at now is a little cheaper yet *normally*).

9-)


As far as large electric bills go - I suspect mine might be larger than yours, but I've got a LOT of machines running, and my other expenses like rent are probably a lot cheaper. Tradeoff, I have to deal with 4 seasons, instead of enjoying San Diego climate.... 8-(

[nuutti]

I have heard from other projects that some users run Prime95 in the background(lowest priority) and other project in the foreground (a little bit higher priority). They want to be sure that if the server of the other project is down and they are not able to get new workunit then they does not waste cpu time. That could explain some of the longest reservations. PC is connected to internet and it reports every 60 days, but practically no progress because almost all cpu time is allocated to other projects.

I Think that slow machines and that kind of users should concentrate to trial factoring.

Yours,

Nuutti

[smh]

I think most of those users know what they are doing, and won't hold up any milestones.

I think the biggest problem are computers on which the client is installed, but which are 'forgotten' or not under control of the person who installed the software. Most of these machines will be slow, and not on 24/7