P-1 benchmarking results in me getting cat 4 assignments

[Uncwilly]

Originally Posted by tuckerkao

It's drkirkby who wants to be in Cat 0 and Cat 1.

Originally Posted by PhilF

why do you care about Cat 0 assignments?

Originally Posted by tuckerkao

I ... unregistered .. server would think my CPU ... thus possible to get Cat 0 assignments regardless the fact I couldn't complete those within a week.

You just said that you are gaming the system to get Cat 0 assignments and then attempt to deflect attention on to someone else.

Have you considered the following scenario (which can happen)?
You run 90% of a test, drop the assignment. PrimeNet reassigns it to the next user asking for an assignment. You then attempt to get it reassigned, but it is assigned to the other user. Would you: a) finish the test anyway and allow the other user to do unneeded work, b) try to contact the user and e-mail them your save files, so they can complete the test, or c) lose all your work on the exponent?

And it have you read the actual rules for qualifying for the different categories?

Computer must have enough LL and DC GHz-days over the last 120 days to indicate the assignment will be completed in 15 days.
Computer must have no expired assignments or bad or suspect results in the last 120 days.
Computer must have returned at least 10 results in the last 120 days.

You said that your machine can't complete the assignment in a week. But if it can in 2 weeks (14 days), it would qualify. You machine has to have 10 results in the last 120 days (averaging 12 days or less per assignment). That is more than a week and and closer to 2 weeks. So, if you are meeting the last condition I copied, you should meet the first. There is no need to game the system if you are doing all of this on a single machine. If you are moving partially completed assignments between machines, that is different.

Why not leave your assignments from PrimeNet alone and see if you don't get Cat 1 assignments (Cat 0 being rare enough that unless you run many machines you might not get any, rather you will have them as Cat 0 when the range moves up to change a Cat 1 to a Cat 0)?

[tuckerkao]

In Category 1, reserve and unreserve are very sensitive, but in high Category 4, almost no other users are working in the same area, thus less than 1/1 million of chance that my exponents may be assigned to another user within minutes if not seconds.

Primenet will never assign a large Category 4 PRP to another user even if all the TFs are up to the GPU72 recommended range and P-1 done.


It seems like drkirkby's CPU will be locked in Cat 4 for at least 120 days due to the suspicious results, if true then he is facing a probable binary choice:

1. Quit GIMPS
2. Test exponents in a range he doesn't want, but has very high demand of needing more PRP tests.

The condition: drkirkby must run a PRP test on an exponent with all the TF and P-1 factoring tasks done by other users.

[tuckerkao]

Originally Posted by Uncwilly

Why not leave your assignments from PrimeNet alone and see if you don't get Cat 1 assignments (Cat 0 being rare enough that unless you run many machines you might not get any, rather you will have them as Cat 0 when the range moves up to change a Cat 1 to a Cat 0)?

Have you considered the following scenario (which can happen)?.

I've unreserved several M105M exponents within 1 hour last week because I don't want to run PRP tests for them, I eventually put 2 larger Cat 4 exponents in my worktodo.txt, then the auto-assigning from the server stopped.

90%+ of my PRP tests locate in a range where no other users even bother to run PRPs for them, thus the scenario will never happen.

I put PRP=N/A,1,2,200204201,-1 in my worktodo.txt

[drkirkby]

Originally Posted by tuckerkao

I've unreserved several M105M exponents within 1 hour last week because I don't want to run PRP tests for them, I eventually put 2 larger Cat 4 exponents in my worktodo.txt, then the auto-assigning from the server stopped.

Since, for reasons I can't possibly comprehend, you want to work in the 168 million range, you could put in prime.txt

MinExponents=168000000
MaxExponents=169000000

Then the server will give you exponents between 168 and 168 million. Anything above 113211942 is category 4 anyway, so you have at least a year to finish them. I don't know when 164 million is likely to be category 1, but I would have thought at least a decade - kriesel has written some estimates.
Can you not see that there are two good reasons not to work with exponents any larger than you can get them?

1. The bigger the exponent, the lower the probability of it being prime.
2. The bigger the exponent, the longer it takes to test.

So it is absurd to work at 168 million rather than 105 million, when the 168 million is going to take you about (168/105)^2.1= 2.7 times as long to test. (I think the power is 2.1 from memory, but it is somewhere around that value).

[drkirkby]

Originally Posted by Uncwilly

Cat 0 being rare enough that unless you run many machines you might not get any, rather you will have them as Cat 0 when the range moves up to change a Cat 1 to a Cat 0)?

I've had four assignments, which were assigned as category 0, on the same machine at the same time. It was not that hard to do, but the difference in time to test a category 0 vs category 1 assignment makes it not worth the bother of chasing the category 0 assignments. But when I did chase them, I got 4 at the same time.

[kriesel]

Originally Posted by drkirkby

I don't know when 164 million is likely to be category 1, but I would have thought at least a decade - kriesel has written some estimates.
Can you not see that there are two good reasons not to work with exponents any larger than you can get them?

1. The bigger the exponent, the lower the probability of it being prime.
2. The bigger the exponent, the longer it takes to test.

So it is absurd to work at 168 million rather than 105 million, when the 168 million is going to take you about (168/105)^2.1= 2.7 times as long to test.

There are a number of sources for the ~2.1 power approximate run time scaling.

1) Low-p empirical benchmarks on a laptop and prime95 (2.094),

2) a fit to theoretical fft length requirement and performance versus exponent over the mersenne.org range (2.117).

3) CUDALucas empirical benchmarking and power fit for the 10⁷ to 10⁸ range on a GTX480 GPU (2.095)

4) Gpuowl v5 10⁸ to 3.32x10⁸ (2.04) https://www.mersenneforum.org/showpo...6&postcount=10

Note, the effect of nonzero fixed startup overhead is to lower the power of the fit.

5) All competitively fast GIMPS PRP, LL or P-1 code is based on the IBDWT, so the order of run-time scaling should be similar, across hardware and software variations. And that's what we find in practice. https://www.mersenneforum.org/showpo...21&postcount=7 (Karatsuba, Toom Cook, and grammar school multiplication have prohibitively steeper run-time scaling.)

Re rate of progress, over the past two years the first-testing wavefront has advanced about 10M/year.
That's more rapidly than the six-year average of 6M/year on which I based a rough projection which implied ~2032 for ~166M. https://www.mersenneforum.org/showpo...5&postcount=11
The rough doubling of GIMPS primality testing throughput by Ben Delo will be absorbed by raising exponent by 2^(1/2.1) ~1.39x.

[tuckerkao]

Originally Posted by kriesel

Re rate of progress, over the past two years the first-testing wavefront has advanced about 10M/year.
That's more rapidly than the six-year average of 6M/year on which I based a rough projection which implied ~2032 for ~166M. https://www.mersenneforum.org/showpo...5&postcount=11
The rough doubling of GIMPS primality testing throughput by Ben Delo will be absorbed by raising exponent by 2^(1/2.1) ~1.39x.

By mid-year 2026, M164M will start to move out from Category 4, then by mid-year 2027, M175M will progress about the same rate.

Users like curtisc will start to receive PRP assignments from those exponents that have been fully TF and P-1 factored which apply to the Category 3 or lower.

[S485122]

Originally Posted by tuckerkao

...
I put PRP=N/A,1,2,200204201,-1 in my worktodo.txt

By setting the AID equal to "N/A" you work outside of PrimeNet. The exponent is not reserved and anybody can claim it unknowingly.
If you leave out that first field and add the "factored to" field : PRP=1,2,200204201,-1,74, then communicate with the server, the exponent
- will be reserved, you will get an AID.
- Prime95 or mprime will not start by trial factoring to from 2²⁸ to 2⁷⁴ before starting the last bit, doubling the trial factoring time, just to reproduce known results.

Your choice for the exponent implies trial factoring from 2⁷⁴ to 2⁷⁵ anyway and this on a CPU !
Before doing P-1 factoring or a PRP test it should be trial factored to 2⁷⁹ on a GPU according the current recommendations.

[tuckerkao]

Originally Posted by S485122

Before doing P-1 factoring or a PRP test it should be trial factored to 2⁷⁹ on a GPU according the current recommendations.

I'd definitely run a P-1 after only 2^78, then bring it up to 2^79 if NF-PM1.

[drkirkby]

Originally Posted by kriesel

Not only your CPU.
a) The incremental or repeated bounds P-1 submissions look like gaming the credits system

Yes, I can understand you could see it as gaming the credits system, but that was not the intention. Common sense would dictate doing P-1 factoring on different exponents would give credit. I really don't care how many total GHz days credit I have - I'm a bit more motivated by credit for just for PRP. If it was real money I might feel a bit different. If I wanted more GIMPS credit, I think trial-factoring with my GPU would be roughly as effective as using one of my Xeon CPUs for P-1 factoring or PRP testing. The GPU would use less electricity than the Xeon. A fairly cheap consumer-grade GPU would give far more credit for trial factoring than my Xeons ever will for P-1 or PRP testing.

Once I did get unearned credit, and I bought that to Georges' attention.

Originally Posted by kriesel

b) Submission for credit is not a necessary part of benchmarking.

I realise that, and I can assure you I am not motivated by GHz days credit.

Originally Posted by kriesel

c) You've previously been advised on how to do benchmarking more productively for GIMPS.

Yes, you said test exponents close to each other, with the same FFT size. I do value your knowledge, but my scientific background tells me to change as little as possible if you want to see the effect of changing one parameter. I knew that if there were any strange results, I would start wondering if it might possibly be because the exponent changed. There were some unexpected results, but at least I am 100% certain they were not due to a change of exponent. Had I got the same results whilst changing both the allocated RAM and exponent, I would have been uncertain if the unexpected result was due to the change of exponent, or change of RAM allocation. For example, generally as RAM was increased, B1 and B2 both increased. But at a large value of RAM (I think either 256 GB or might have been larger), B1 fell. Had the exponent changed, I would have probably suspected that the cause. But since the exponent was identical, I know that was not the cause. So I don't regret testing the same exponent, although I had not foreseen this consequence. But it all became totally irrelevant after one post from chalsall
https://www.mersenneforum.org/showpo...&postcount=368
which was seen by Prime95, but not commented on.

Originally Posted by kriesel

d) Creating an entire new thread about a single question is something various folks have advised against before, including moderators who can threaten or take certain actions on a per-forum-user basis, yet here it is again.

The way the threads work on here is one of the reasons I've decided to quit work on GIMPS. I just can't hack this thing where everything goes in one long thread. In every other mailing list or forum I have used, that would be seen as hijacking a thread. But here the norm seems to be to have a huge thread of vaguely related posts. You wrote one thread had nearly 3500 posts in it. To me that is silly, and one of the reasons I'm quitting working on the GIMPS project. The forums are the worst run of any I have ever come across. Mathematica is used by many mathematicians, but the Mathematica forums are much better run than here, and there are few if any arrogant people. (I've not come across anyone other than helpful people, but I'm sure there must be some arrogant ones there too.)

Originally Posted by kriesel

e) A little thought would have provided an answer to producing the effect you asked about.

It could be one of 3 reasons, although most likely seems to be sending the exact same results more than once. I repeated a couple of tests where the results seemed a little odd. The server reported it had them. I don't think it was because I submitted multiple redundant results, as I think I got category 1 assignments after that. I think the problem was most likely submitting the same result twice, but I am not sure.

Originally Posted by kriesel

f) You neglect to mention the application or version involved. One in the know can deduce it or narrow it down from other post content, but requiring that of each reader is impolite at best.

I take your point. It is annoying when people mention software, but don't say the version.

Originally Posted by kriesel

g) Proofread, spell check, grammar check, read for adequate content. The little red wavy underline is browser spell-check. Use it. Always. ("bencharking")

It probably does not help the fact that benchmarking is also seen as a spelling error.

Originally Posted by kriesel

h) There's an avatar and account info displayed at left of each post. Signatures in post bodies are conspicuously absent in almost all forum posts. Signing posts is redundant and wasteful use of forum storage space. Signing uses relatively few bytes per post, but if every post by everyone was redundantly signed, it would add up to quite a lot.

This is another reason I've decided to leave the project - the comments about trivia such as storage space of a name. Yet the other day I see you requesting the storage space for zip files was increased from 4 MB to 8 MB. A typical LZMA compressed file will be 15% smaller than a zip one. So why use 100's KB using archaic compression algorithms rather than more more modern alternatives, then worry about 4 bytes of a name? I have never come across any other forum where people object if you stick your name. I try to remember not to do it, as Uncwilly asked me not to, but I do occasionally forget. The pettiness just gets on my tits.

Originally Posted by kriesel

i) You might feel that you're being singled out, or picked at for minor things. Things that might be let slide, if they were alone, may not be when someone ignores or actively resists guidance or advice on several fronts and exhausts another's patience. (A certain leeway may be given the major players, such as forum owner, authors of major software, server admins, etc, in various matters, that would not be given us mere users.)

Yes, sometimes I do feel it is minor things - like complaining since someone put their name. But the main reason I'm quitting GIMPS work is that there are some pretty damn nasty people on here, and George is aware of it, but ignores it. In particular chalsall is very arrogant. Apparently he likes to insult lots of people, although ignores those who have been here a long time. I'm led to believe he gets away with it because of his GPU to 72 project.

Uncwilly is also very arrogant. His comment about a joke I made in the Lounge

https://www.mersenneforum.org/showpo...&postcount=878
was particularly childish. He is a moderator here, and just irritates me.

You can sometimes appear a bit arrogant yourself, but at least you generally try to be helpful. But overall, I think there's a pretty nasty atmosphere in these forums, and it is not all directed at me. I've received several PMs from people who have been here years and comment on the tone of the forums. One commented Just ignore chalsall, he's an arrogant asshole that likes to insult people.

Originally Posted by kriesel

You may note I have not answered the question of how to prevent a worker from interacting with the server, or how to beat the 120 days penalty imposed by the server. That is to avoid enabling or rewarding bad behavior.

I'm not 100% sure why you feel the need to not tell me how to avoid a worker interacting with the server, but that's your choice. Probably because you think it is better the worker does interact with the server. Anyway, I am past caring. I've put

DaysOfWork=0

in prime.txt of my main workstation and on another server I use. I'll finish those exponents. I did contemplate getting a double-check of M105212323, as I submitted that without the proof being generated for some unknown reason, but have decided not to bother. I will terminate a couple of Amazon instances, and unreserve the exponents. I can assure you, if I wanted to get around the 120 day penalty I'm quite capable of working out a way around it. Maybe the best way is to simply ask the database admin - he seems a pleasant guy, but I know how to move assignments between machines. I put category 1 exponents on a server that can only get category 3, but can complete category 1 exponents in about 25 days. But as I say, I'm not going to bother. Uncwilly, chalsall and a few others have convinced me it is not worth staying. And despite some well meaning PMs from several people, I'm not going to hang around much longer.

[kriesel]

Intended as helpful:

Forum participation and GIMPS participation are two different things. One can quit mersenne forum participation and continue GIMPS computation.

For words we're sure are spelled correctly, in most browsers, right click, add to dictionary, handles them permanently. And there are online dictionaries for confirmation, and browser dictionary undo.

Today's forum statistics show 520,736 posts in 20,109 threads; 4240 members. (Bottom of https://mersenneforum.org/index.php)
If all 520736 posts had 6-char signatures, a rough minimum when including a line feed before "David" or whatever, that's a few MB; 20-char, ~10MB. I had not done the math earlier. Space is cheap, until it nudges pricing over a threshold.
IIRC: Attachments are handled differently than post text. Post text increases the forum database size.
Attachments are separate files.
Attachments >4MB are only of interest to me in helping solve problems or perceived problems, such as absence of python on Windows installs, or some users' aversion or possible policy obstacles to Python.
Most attachments would still be small compared to 1MB. (My pdf attachments are frequently <50KB)

One of the reasons I was given a blog was so reference info could be organized in one location that becomes known, and get updated over time, rather than inflating existing large threads with re-posts for updates.
Attachments there get removed and replaced as part of the update process. And potentially are used by many.

In a group of over 4000 members, it is not surprising there are some "characters". And some are more reliably civil than others.

I think it would be a poor choice to post publicly, methods for defeating the assignment rules or operation of any rules. That's not personal.

There is the possibility of adding one or more selected ordinary forum users to your own mersenne forum "ignore list". This is not however effective in blocking posts or PMs or other actions from moderators that one may feel are being rude or abusive or otherwise inappropriate. I have been recommended the course of appealing to multiple moderators in that case. I suggest reserving that for only the clearest cases (smoking gun with video equivalent), as in other cases it may result instead in additional unwanted attention from multiple moderators supporting the moderator's action. Also choosing well which moderators to cc. That said, I don't envy the moderator task one bit.

One of the things I sometimes pull back and consider is, why and how much should we care what some user or moderator thinks, that we've never met in real life and probably never will. Not saying others' feelings don't matter, but remembering to keep things in perspective. It's also difficult to divine whether a person is posting from arrogance, ignorance, irritation, or something else, and easy to mistake the mindset of the poster as something other than what was intended. Moderators being human are not immune to this error.
Striking a balance between being open to coaching and not tolerating abuse is a judgment call.