Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
1)
Message boards :
Problems and Help :
Newbie: Deadline warning
(Message 162864)
Posted 3 days ago by  pschoefer
PrimeGrid 4.02 Do You Feel Lucky? (OCLcudaGFNEXTREME) genefer_extreme_112422471_1 02d,04:12:37 (00:12:12) 0.39 37.832 614d,22:32:45 15d,12:50:38 Running High P., Deadline warning 66.7 °C 99 Slim-7i - Win 0.112C + 1NV
If I'm reading this correctly, the task's progress is at 37.832%. PrimeGrid's applications report their progress quite accurately, so a much better estimate for the remaining time would be ~86 hours (actually, PrimeGrid tells BOINC to estimate the remaining time this way, but BOINC sometimes messes this up, e.g., after a reboot).
So, as long as your computer is not slowing down considerably, it should have no problem finishing this task in time.
|
2)
Message boards :
Number crunching :
International Women's Day Challenge
(Message 161026)
Posted 81 days ago by pschoefer
I think there is a minor bug in the challenge stats. If I divide the score by the number of tasks for each participant, I get values between 39.91357038 and 39.91357125, i.e. no difference up to the 6th decimal place. However, as the credit granted for a SGS task resulting in a prime is twice as much as for a normal task, the variation should be much larger. Even for Nick's 311923 tasks, an additional prime task would change the 4th decimal place in the credit per task ratio.
|
3)
Message boards :
Number crunching :
Tour de Primes 2023
(Message 160929)
Posted 85 days ago by pschoefer
The same information in graphical format:
|
4)
Message boards :
Generalized Fermat Prime Search :
Do You Feel Lucky? Jumped from 2+ days to 6+ days remaining. Why?
(Message 160901)
Posted 87 days ago by pschoefer
Remaining time is always incorrect with Boinc. It evaluates complex statistics and generates a wrong estimate.
The simple formula elapsed = progress * total and then remaining = elapsed / progress - elapsed is accurate.
I don't know why Boinc doesn't use this relation and if it will be fixed one of these days.
BOINC does not use this relation by default, because some applications have progress bars that are inaccurate or do not update very often, or have no progress bar at all. However, both the project and the user (via app_config.xml) already can tell the client to use the simple linear interpolation (fraction_done_exact). PrimeGrid actually does this, but it looks like it does not work after a restart.
|
5)
Message boards :
Number crunching :
Tour de Primes 2023
(Message 160825)
Posted 92 days ago by pschoefer
After 26 days of crunching, I still can't get a prime this year, sadly :( My workload is basically 45 PPS WUs and 600 GFN16 WUs per day, which is actually the same as what it was in 2022.
I think TdP is more and more difficult for students like me to score a prime since PPSE is no longer counted as top 5000 and small verification WU makes everyone 100% which means those super PG crunchers save a significant amount of time from not doing 2nd tasks
For example, in one day, PG heroes like tng & Scoutt run 10k first tasks and 3k 2nd tasks (77% first-rate) while I run 400 first tasks and 200 2nd tasks (67% first-rate, which is approximately true) in 2022 but now they run 13k first tasks while I run 600 first tasks, so obviously they can get 3k more tasks while I only get 200 more in one day in 2023 now. Is it correct to do math like this lol?
It's true that the difficulty to find a Top5k prime has slightly gone up, but the jump from PPSE last year to GFN-16 this year is not that big and, for most computers, compensated for by the fast DC mechanism. There might be a much bigger difficulty jump next year, as GFN-16 will almost certainly be too small, leaving PPS-MEGA as the smallest option (unless a new subproject between those two is created).
While your calculation is correct, your conclusion is not. Yes, the big guys will do more "1st" tasks and therefore find more primes... but so do you and everyone else! This is part of the reason why in total already more than twice as many primes as last year have been found.
Note that, in your example, you profit more than the big guys from the fast DC mechanism: If the big guys did 10k "1st" tasks per day and you did 400 "1st" tasks per day, they would on average find 25x as many primes as you. If they now do 13k tasks per day and you do 600 per day, they will on average find only 21.67x as many primes.
|
6)
Message boards :
Number crunching :
Tour de Primes 2023
(Message 160796)
Posted 93 days ago by pschoefer
As the TDP finishes soon, I was wondering which project should I run to maximize my chances to discover a prime (hopefully a MEGA one) before the end. Obviously GFN-16 for GPU, but for the CPU is it PPS-MEGA or GFN-16?
On average, you'll have to process roughly twice as many PPS-MEGA tasks than GFN-16 tasks to find a prime. So GFN-16 very clearly maximizes your chances, given that the tasks are also shorter. It obviously won't be a mega prime, though.
Also, I am running a single core per task using an Intel Xeon E5-2697 V4, which has 18 cores and 45 MB of L3 cache. Am I doing this right? Or should I multithread?
Cache is sufficient for one task per core, so there would be no benefit from multithreading. For GFN-16, it might even slightly increase the throughput if you enabled hyperthreading and ran 36 tasks in parallel.
|
7)
Message boards :
Number crunching :
Tour de Primes 2023
(Message 160776)
Posted 94 days ago by pschoefer
I wonder if anyone will find another GFN 18 to pip tng at the finish line.
Using Yves's formula from earlier in this thread, we can expect one GFN-18 prime per ~52k candidates right now. The number of completed tasks per day has been oscillating around 20k since the beginning of the month according to this graph. I've been told that the fast proof tasks are not included in that graph, so it's one task per candidate and we can expect one prime every 2.5 days (i.e., one just a few hours ago and another one on Monday... yes, I know that it doesn't work that way :)).
Let's see if the pattern continues that each GFN-18 prime is found by a different person. I only have my Raspberry Pi 3B+ working on that subproject (~32 hours per task using all four cores), so don't bet on me. :D
|
8)
Message boards :
Number crunching :
Tour de Primes 2023
(Message 160748)
Posted 95 days ago by pschoefer
Hmm for give me for saying pschoefer but are you sure about that statement? Most found on the 19th weren't via cloud services?
There must not have been much in it if true.
Of the 78 primes that are visible until now:
- 23 have been found using a mainstream CPU (12 Intel Core ix, 11 AMD Ryzen)
- 19 have been found using a mainstream GPU (18 NVIDIA GeForce, 1 AMD Radeon)
- 4 have been found with Intel Xeon CPUs, but those computers have GPUs or operating systems that I doubt any cloud provider would dare to offer
- 28 have been found by an obvious cloud computer (18 DigitalOcean, 8 Google, 1 Amazon, 1 Microsoft)
- 4 have been found with server CPUs that are extremely uncommon in home or business settings, so they might be from a lesser known cloud service or dedicated server rentals
That's at least 46 of 78 primes found on computers that look "ground-based" rather than cloud-based to me, pretty much a reversal of last year's Mountain Stage (12 "ground primes" vs 17 "cloud primes").
There are two caveats with these numbers:
1. The results are not final yet, so there may be some hidden primes. I see no reason why they should heavily skew to the cloud, though.
2. It's not impossible to fake the CPU/GPU/OS information that the clients report to the server. However, I can't think of any reason why a "ground user" would make their computers appear like cloud computers, and the cloud users don't appear to feel uncomfortable using the cloud, so it's unlikely that this is happening here. In addition, the dominance of DigitalOcean and Google among the "cloud primes" suggests that most cloud users are buying their cloud instances through The Science Cloud, so they may not be able to fake the client information even if they wanted to.
So, all in all, I'm fairly sure about my statement.
|
9)
Message boards :
Number crunching :
Tour de Primes 2023
(Message 160731)
Posted 95 days ago by pschoefer
The amount of GFN-16 being done continues to be unprecedented. The first GFN-16 prime of TdP 2023 had a b of 186 million. We're currently above 241 million. That's an increase of 55 million in less than 23 days. By contrast, it took 10 years to do the first 55 million in the project. In fact, we were at b=55 million only 3 years ago
I guess that's thanks to cloud computing.
There is this perception out there that 99.999999999999% of our computing is being done by bogymen running phantom virtual cloud computers in orbit around Venus. Or something just as silly.
Yes, cloud computing is a thing, and for ONE DAY (the Mountain Stage) it was actually a big thing, but most of the time it's a relatively small presence. People aren't en-masse deciding to take out a second mortgage to either buy more computers, or to rent a bunch of computers. Certainly not for a whole month. For a day, it's fun. For a month, it's real money.
To help set perceptions straight, it should also be noted that even on that one day the majority of primes was not found by cloud computers - unlike in last year's Mountain Stage!
In addition to the advances in software and hardware leading to the massive leading edge advance this month, GFN-16 was also just not all that "interesting" in the past: Yes, it has been the most promising option for finding a Top5k prime on a GPU for several years, but there were much more attractive options on CPU until recently (PPSE until last year's TdP, PPS and MEGA until fast DC was introduced for GFN a few months ago). In all the years since it was migrated from PRPNet to BOINC, there was only one GFN-16 challenge (Pierre de Fermat's Birthday Challenge in November 2017) and that was together with GFN-15 and GFN-17-Low (which, according to my notes, were more efficient in terms of points per run time during that challenge). Finally, all GFN subprojects count towards the same badge, so the badge hunters also had no particular reason to crunch GFN-16.
|
10)
Message boards :
Number crunching :
A Prime Chinese New Year Challenge
(Message 159648)
Posted 124 days ago by pschoefer
Doing some pre-testing on GFN-17. Seeing following on CPU tasks:
Warning: b > 125,000,000: the test may fail. Anything to be worried about? No problems on the handful of units I've run so far.
At b = 125M, genefer is starting to compute and print the round-off error. On your computer, it is <= 0.375 (it is ~ 0.4 without fma).
The real limit where some CPU errors are going to be reported should be greater than 140M. For the moment, it is just monitoring.
It looks like there are a few unlucky numbers at b<140M. During the challenge, I had two tasks ending with "validation failed!": b=135028388 using avx and b=136030622 using sse4. In both cases, I was able to reproduce this on a different computer with avx, while fma produced a good result, so I'm quite confident that it's the b limit and not some random hardware error.
|
Next 10 posts
|
