|
2019-01-11, 17:23
|
#1816 | |
|
Serpentine Vermin Jar
Jul 2014
2·13·131 Posts |
Quote:
The "trick" of starting an unassigned run on an exponent and forcing a check-in will still work since the server says "ah, well, this machine did start, I may as well assign it if available". It's a nicety, but cumbersome for GPUs. |
|
|
|
2019-01-11, 19:03
|
#1817 | |
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
172158 Posts |
Quote:
DC, clearly states, cat 2 3 and 4 are eligible for manual assignment. Cat 2 places a condition on it: "If manual testing, user must have signed up for the smallest exponents (see setting in green above)" LL test, same. category bounds change over time. currently: cat 0 <82864802 cat 1 <84858848 (difference from cat 0, 1,994,046) cat 2 <86645330 (difference from cat 1, 1,786,482) cat 3 <88493310 (difference from cat 2, 1,847,980) cat 4 everything else in mersenne.org's p<109 range. See also the thresholds for system speed etc. Simple test: obtain one smallest available first time LL assignment. I got Test=(aid redacted),83104493,76,1 That's a cat 1 exponent manually assigned, via the https://www.mersenne.org/manual_assignment/ page. Cat 1 deadlines: "Assignments are recycled if assignment is not started within 20 days or does not report in for 30 days or when assignment is more than 90 days old." Cat 2 deadlines: "Assignments are recycled if assignment is not started within 30 (unless manual testing) days or when the exponent moves midway into the first category and the assignment is more than 180 days old." A manual assignment put on a gpu will appear to the server to be unstarted, during its entire computation, since there's no provision for reporting percent complete updates manually from a gpu application. So a cat 1 manual assignment should be completed and reported within 20 days of its assignment, to avoid assignment expiration and the appearance of very fast poaching of the assignment then reassigned to someone else. (So into the short queue on a fast gpu that manual assignment test exponent above goes. If put on my slowest discrete gpu, Quadro 2000, it would take around 38 days. I paradoxically need to reserve bigger slower exponents to have enough time to complete them before they expire; cat 3 or 4 or high cat 2 when reserved. And the category boundaries move noticeably during the run times. An 88M exponent takes about 43 days on a Quadro 2000, in which time the category boundaries can advance by about 708,500. That's 38% of the average current width1,876,169.3 of categories 1, 2, and 3.) Last fiddled with by kriesel on 2019-01-11 at 19:10 |
|
|
|
|
2019-01-12, 18:03
|
#1818 | |
|
Serpentine Vermin Jar
Jul 2014
2·13·131 Posts |
Quote:
I even briefly tinkered with the idea myself, when I was regularly getting large batches of manual assignments of my own for the strategic DC work. I did it all with manual assignments, of course, since it was a very select list of exponents, and I would sometimes queue of several weeks worth. I didn't want the client to automatically communicate because it might pick up assignments I didn't want, or expire some, etc. So I started some scripts that would reach out to each worker, read the backup file and get the current iteration, etc. It would have had to look in the worktodo as well to get the assignment ID for that exponent. Anyway, in theory it would have all worked that way. I got as far as figuring out what the POST should look like (and I think there was a requirement that the "client" needs to update the CPU info first before doing any progress updates). Why didn't I finish that up? Well, because I realized that since I have access to the server itself, I could take that same info and output a script to update my progress directly in the database. LOL ... So yeah, I took the easy way. But yeah, someone who reads the API, and maybe monitors what a real client is sending just to confirm how it's working, should be able to do it. There are other projects out there like GPU72 that currently get assignments using the web pages, and it would be so much easier (after the initial work) to use the API, but oh well... we tend to stick with what works, I suppose. |
|
|
|
|
2019-01-12, 18:56
|
#1820 | |
|
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
172158 Posts |
Quote:
I can't find it right now, using the forum search tool, but I wrote a while back about the disparity between the cpu-oriented data the Primenet API seeks and the different memory model gpus have. It looked to me a bit like trying to fit one's hand in a shoe, or wear a glove on one's foot. (There does not seem to be a capability in the forum search tool, to specify keword1 AND keyword2. I get inclusive or results for searches. Also API is not a keyword with any hits at all. Inclusive-or generates the apparently maximal matches returned too easily, not searching back far enough.) Oh, here it is: a suggestion for extension of the API so it fits gpu computing. https://www.mersenneforum.org/showpo...&postcount=427 If we give the Primenet front end gpu parameters fit as well as practical to the cpu-oriented parameters it wants, will it accept and handle them sensibly in the back end? For example, for number of cpu cores, do we use 1, # of ROPS, # of TMU, # of shaders? Will Primenet accept such large numbers? Will it expect to potentially be able to run an independent assignment (worker) on each? For the GTX1080, 2560 cores per the NVIDIA specifications at ("full specs" link in) https://www.nvidia.com/en-us/geforce...orce-gtx-1080/ How best to map the memory parameters? Last fiddled with by kriesel on 2019-01-12 at 19:25 Reason: found my old post (with keyword extension) |
|
|
|
|
2019-02-04, 00:48
|
#1823 |
|
Serpentine Vermin Jar
Jul 2014
65168 Posts |
New list
Recently I've been doing a deep look at error codes and how they affect the overall good/bad status.
I'm not sure why (and there's probably some old reason), but George has set things so that any # of "illegal sumout" errors are *not* treated as "suspicious". However, when I look at the historical end status of such results (good or bad), I've seen that if there is even a single "illegal sumout" recorded, the odds of it being bad are about 37.5%. Strange thing is that it doesn't seem to matter much if you have just one, or a lot of them (over 10). The overall odds of being bad are roughly the same. Like, for those with over 20 of them, the odds only go to ~ 40% chance of being bad. "illegal sumouts" are recorded in the 2nd most significant byte of the error code, so, for example an error code like "00230000" the illegal sumouts are "0x23" or 35 decimal. In versions of Prime95 >= 29.3 it's actually only the lowest nibble of that byte, so it'd only be "0x3" in that example. The upper nibble of that byte is now set aside for a count of PRP errors. Meaning new versions would only max out at being able to record up to 15 such errors. But like I mentioned, it doesn't matter if it's one or 20+, the odds of being bad are roughly the same. Anyway, right now I have a list of 74 exponents below 70M with at least one illegal sumout during its run. Here's that list, sorted by exponent. I included the decimal count of illegal sumouts (column "is") in case anyone is curious. Code:
exponent is worktodo 46927799 1 DoubleCheck=46927799,73,1 47316499 12 DoubleCheck=47316499,73,1 47371187 1 DoubleCheck=47371187,73,1 47715463 1 DoubleCheck=47715463,73,1 47731771 2 DoubleCheck=47731771,73,1 47777167 1 DoubleCheck=47777167,73,1 47905967 1 DoubleCheck=47905967,73,1 47961779 1 DoubleCheck=47961779,73,1 48302941 1 DoubleCheck=48302941,73,1 48418189 1 DoubleCheck=48418189,73,1 48461863 1 DoubleCheck=48461863,73,1 49287883 1 DoubleCheck=49287883,73,1 49486603 1 DoubleCheck=49486603,73,1 49743829 1 DoubleCheck=49743829,73,1 50334617 32 DoubleCheck=50334617,73,1 50412689 1 DoubleCheck=50412689,73,1 50539561 2 DoubleCheck=50539561,73,1 50567743 2 DoubleCheck=50567743,73,1 50567893 1 DoubleCheck=50567893,73,1 50970839 1 DoubleCheck=50970839,73,1 50988037 1 DoubleCheck=50988037,73,1 51813929 1 DoubleCheck=51813929,73,1 52875653 1 DoubleCheck=52875653,73,1 54794357 1 DoubleCheck=54794357,73,1 54963851 1 DoubleCheck=54963851,73,1 55473161 1 DoubleCheck=55473161,74,1 55473541 1 DoubleCheck=55473541,74,1 55553083 1 DoubleCheck=55553083,73,1 56604599 1 DoubleCheck=56604599,73,1 56636033 2 DoubleCheck=56636033,73,1 56648089 1 DoubleCheck=56648089,73,1 56768977 1 DoubleCheck=56768977,73,1 56893553 1 DoubleCheck=56893553,73,1 57994177 1 DoubleCheck=57994177,74,1 59059157 1 DoubleCheck=59059157,75,1 60724151 1 DoubleCheck=60724151,74,1 61062347 1 DoubleCheck=61062347,74,1 61318423 1 DoubleCheck=61318423,74,1 61419041 1 DoubleCheck=61419041,74,1 61444667 1 DoubleCheck=61444667,74,1 61545647 1 DoubleCheck=61545647,74,1 62303069 1 DoubleCheck=62303069,74,1 62452219 1 DoubleCheck=62452219,74,1 62493631 1 DoubleCheck=62493631,74,1 62871191 1 DoubleCheck=62871191,74,1 62917997 2 DoubleCheck=62917997,74,1 63955781 1 DoubleCheck=63955781,74,1 64333307 1 DoubleCheck=64333307,74,1 64381453 1 DoubleCheck=64381453,74,1 65361887 1 DoubleCheck=65361887,74,1 65392891 1 DoubleCheck=65392891,74,1 66358169 6 DoubleCheck=66358169,74,1 66407041 1 DoubleCheck=66407041,75,1 66641867 1 DoubleCheck=66641867,75,1 67124983 1 DoubleCheck=67124983,75,1 67146199 1 DoubleCheck=67146199,75,1 67265641 1 DoubleCheck=67265641,75,1 67316213 1 DoubleCheck=67316213,75,1 67457959 1 DoubleCheck=67457959,75,1 67458731 1 DoubleCheck=67458731,75,1 68255843 1 DoubleCheck=68255843,75,1 68592781 1 DoubleCheck=68592781,75,1 68776811 1 DoubleCheck=68776811,75,1 68786083 1 DoubleCheck=68786083,75,1 68852843 1 DoubleCheck=68852843,75,1 68955583 1 DoubleCheck=68955583,75,1 69193373 1 DoubleCheck=69193373,75,1 69267971 1 DoubleCheck=69267971,75,1 69386683 1 DoubleCheck=69386683,75,1 69433369 4 DoubleCheck=69433369,74,1 69445963 1 DoubleCheck=69445963,75,1 69479407 1 DoubleCheck=69479407,75,1 69479569 1 DoubleCheck=69479569,75,1 69717211 1 DoubleCheck=69717211,75,1 |
|
|
|
2019-02-04, 01:28
|
#1824 |
|
P90 years forever!
Aug 2002
Yeehaw, FL
17×487 Posts |
I took the first one:
Code:
DoubleCheck=46927799,73,1 |
|
|
|
2019-02-04, 10:52
|
#1825 | |
|
Banned
"Luigi"
Aug 2002
Team Italia
114018 Posts |
Quote:
Code:
47316499 12 DoubleCheck=47316499,73,1 47371187 1 DoubleCheck=47371187,73,1 47715463 1 DoubleCheck=47715463,73,1 47731771 2 DoubleCheck=47731771,73,1 47777167 1 DoubleCheck=47777167,73,1 47905967 1 DoubleCheck=47905967,73,1 |
|
|
|
|
2019-02-04, 15:18
|
#1826 |
|
6809 > 6502
"""""""""""""""""""
Aug 2003
101×103 Posts
2B3B16 Posts |
I took some. Here is the remaining list.
Code:
50334617 32 DoubleCheck=50334617,73,1 50412689 1 DoubleCheck=50412689,73,1 50539561 2 DoubleCheck=50539561,73,1 50567743 2 DoubleCheck=50567743,73,1 50567893 1 DoubleCheck=50567893,73,1 50970839 1 DoubleCheck=50970839,73,1 50988037 1 DoubleCheck=50988037,73,1 51813929 1 DoubleCheck=51813929,73,1 52875653 1 DoubleCheck=52875653,73,1 54794357 1 DoubleCheck=54794357,73,1 54963851 1 DoubleCheck=54963851,73,1 55473161 1 DoubleCheck=55473161,74,1 55473541 1 DoubleCheck=55473541,74,1 55553083 1 DoubleCheck=55553083,73,1 56604599 1 DoubleCheck=56604599,73,1 56636033 2 DoubleCheck=56636033,73,1 56648089 1 DoubleCheck=56648089,73,1 56768977 1 DoubleCheck=56768977,73,1 56893553 1 DoubleCheck=56893553,73,1 57994177 1 DoubleCheck=57994177,74,1 59059157 1 DoubleCheck=59059157,75,1 60724151 1 DoubleCheck=60724151,74,1 61062347 1 DoubleCheck=61062347,74,1 61318423 1 DoubleCheck=61318423,74,1 61419041 1 DoubleCheck=61419041,74,1 61444667 1 DoubleCheck=61444667,74,1 61545647 1 DoubleCheck=61545647,74,1 62303069 1 DoubleCheck=62303069,74,1 62452219 1 DoubleCheck=62452219,74,1 62493631 1 DoubleCheck=62493631,74,1 62871191 1 DoubleCheck=62871191,74,1 62917997 2 DoubleCheck=62917997,74,1 63955781 1 DoubleCheck=63955781,74,1 64333307 1 DoubleCheck=64333307,74,1 64381453 1 DoubleCheck=64381453,74,1 65361887 1 DoubleCheck=65361887,74,1 65392891 1 DoubleCheck=65392891,74,1 66358169 6 DoubleCheck=66358169,74,1 66407041 1 DoubleCheck=66407041,75,1 66641867 1 DoubleCheck=66641867,75,1 67124983 1 DoubleCheck=67124983,75,1 67146199 1 DoubleCheck=67146199,75,1 67265641 1 DoubleCheck=67265641,75,1 67316213 1 DoubleCheck=67316213,75,1 67457959 1 DoubleCheck=67457959,75,1 67458731 1 DoubleCheck=67458731,75,1 68255843 1 DoubleCheck=68255843,75,1 68592781 1 DoubleCheck=68592781,75,1 68776811 1 DoubleCheck=68776811,75,1 68786083 1 DoubleCheck=68786083,75,1 68852843 1 DoubleCheck=68852843,75,1 68955583 1 DoubleCheck=68955583,75,1 69193373 1 DoubleCheck=69193373,75,1 69267971 1 DoubleCheck=69267971,75,1 69386683 1 DoubleCheck=69386683,75,1 69433369 4 DoubleCheck=69433369,74,1 69445963 1 DoubleCheck=69445963,75,1 69479407 1 DoubleCheck=69479407,75,1 69479569 1 DoubleCheck=69479569,75,1 69717211 1 DoubleCheck=69717211,75,1 |
|
 |
Similar Threads
|
||||
| Thread | Thread Starter | Forum | Replies | Last Post |
| Double-Double Arithmetic | Mysticial | Software | 52 | 2021-04-23 06:51 |
| Clicking an exponent leads to 404 page | marigonzes | Information & Answers | 2 | 2017-02-14 16:56 |
| x.265 half the size, double the computation; so if you double again? 1/4th? | jasong | jasong | 7 | 2015-08-17 10:56 |
| What about double-checking TF/P-1? | 137ben | PrimeNet | 6 | 2012-03-13 04:01 |
| Double the area, Double the volume. | Uncwilly | Puzzles | 8 | 2006-07-03 16:02 |
