mersenneforum.org drafts in progress, comments would probably be premature
 Register FAQ Search Today's Posts Mark Forums Read

 2020-10-24, 20:42 #12 kriesel     "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 22·3·52·17 Posts GPU models no longer available to me for test GTX 480 Tesla C2075 GTX 1070 GTX 1080Ti Last fiddled with by kriesel on 2021-01-29 at 17:08
 2020-12-08, 22:29 #13 kriesel     "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 117548 Posts Software etc for numbers of other forms Someday, maybe, summarize software, computation type, hardware type, number name, formula, useful threads etc for non-Mersenne-numbers https://www.mersenneforum.org/showthread.php?t=25874 Wagstaff Riesel Proth Fermat Last fiddled with by kriesel on 2020-12-08 at 22:46
 2021-01-13, 16:34 #15 kriesel     "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 117548 Posts Windows Subsystem for Linux (WSL) early draft, maybe becomes a whole thread Requirements: 64-bit hardware supported by Windows WSL1: Windows 10 x64 v1607 or later per https://www.computerhope.com/issues/ch001879.htm WSL2: Windows10 x64 ~v1909 or later, and a processor that supports vritualization (VT-X?), from personal experience. Checking whether WSL is installed and what version: in a Windows command prompt, "wsl -l -v" https://onecompiler.com/questions/3w...wsl-versionupd And it can be useful to "title wsl check" and leave the small command prompt box open as a reminder. Code:  NAME STATE VERSION * Ubuntu Stopped 1 Unless there is WSL but no Linux distro installed: Code: C:\users\User\Documents>wsl -l -v Windows Subsystem for Linux has no installed distributions. Distributions can be installed by visiting the Microsoft Store: https://aka.ms/wslstore Or if WSL is not installed either: Code: 'wsl' is not recognized as an internal or external command, operable program or batch file. Differences between WSL1 and WSL2 https://docs.microsoft.com/en-us/win...mpare-versions Installing WSL2: update to required Windows 10 version or higher. For x64 systems: Version 1903 or higher, with Build 18362 or higher. For ARM64 systems: Version 2004 or higher, with Build 19041 or higher. Install WSL 1. Then update to WSL 2. https://www.windowscentral.com/how-i...sl2-windows-10 Simplified WSL2 install: Windows 10 build 20262 or higher https://docs.microsoft.com/en-us/win.../install-win10 A handy reference, including adding GUI and RDP to WSL https://adamtheautomator.com/windows...tem-for-linux/ Top of reference tree: https://www.mersenneforum.org/showpo...22&postcount=1 Attached Thumbnails   Last fiddled with by kriesel on 2021-02-02 at 18:38
 2021-01-23, 00:02 #16 kriesel     "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 117548 Posts Mersenne semiprimes Semiprimes are natural numbers that are products of exactly two factors other than one and themselves (and they may be squares). https://prime-numbers.info/article/semiprimes http://oeis.org/A092561 lists Mersenne numbers that are semiprimes without regard to whether n is composite. n Mn 4 15, 9 511, 11 2047, 23 8388607, 37 137438953471, 41 2199023255551, 49 562949953421311, 59 576460752303423487, 67 147573952589676412927, 83 9671406556917033397649407, 97 158456325028528675187087900671, 101 2535301200456458802993406410751 Subsetting to prime exponent < 100 gives p for Mp and factors 11 2047 f=23; 89 23 8,388,607 f=47; 178481 37 137438953471 f=223; 616318177 41 2,199,023,255,551 f=13367, 164,511,353 59 576,460,752,303,423,487 f= 179951, 3 203431 780337 67 147,573,952,589,676,412,927 = 193 707721 × 761838 257287 83 9,671,406,556,917,033,397,649,407 = 167 × 57912 614113 275649 087721 97 158,456,325,028,528,675,187,087,900,671 = 11447 × 13 842607 235828 485645 766393 below p=100, 8 Mersenne semiprimes of prime exponent vs. 10 Mersenne primes. in p= 100 to 200, 8 semiprimes vs. only 2 primes of prime exponent. in p=200 to 500, 9 vs 0. https://www.mersenne.org/report_expo..._hi=500&full=1 Last fiddled with by kriesel on 2021-01-29 at 17:22 Reason: add a semiprime definition and link
2021-03-25, 16:19   #17
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

22×3×52×17 Posts
CUDALucas export

The zip file contains a small plain text documentation file, Perl source code, and a Windows executable.
The program accepts up to 3 command line parameters at a time, although there is not much point to more than two.

Parameter choices are
• a filename in the current directory that is a CUDALucas checkpoint or save file, as the first command line parameter
• option -ASCII, for ASCII hex encoding of bulk residue data
• option -headeronly, for producing only the header section. This outputs to filename.MNE.hdr.txt and to console. Useful for obtaining iteration number, orffset, Res64, fft length without generating a big export file.
• default is binary output of bulk residue data after the header section, if neither -ASCII nor -headeronly are specified
• if no filename is provided, the program will ask for one.
This program is to implement CUDALucas export for V2.06 or V2.05x, per the Mersenne Neutral Exchange Format V1.0 draft 6.
Fastest and most compact is -headeronly.
Slowest by far and largest output is -ASCII.
Attached Files
 cudalucasexport.zip (1.26 MB, 14 views) Mersenne neutral exchange format draft 6.pdf (136.1 KB, 15 views)

Last fiddled with by kriesel on 2021-03-28 at 23:19

 2021-03-26, 10:31 #18 kriesel     "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 22×3×52×17 Posts PRP types
2021-03-29, 02:00   #19
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

22×3×52×17 Posts
unreliable gpu case study

The gpu in question is an XFX Radeon VII with 16GB of Samsung ram. It is significantly more error-prone, as indicated by GpuOwl GEC error detections, than the Hynix-ram Radeon VIIs, even at nominal gpu ram clock rate or below. The Hynix ram gpus can go weeks between errors at over 1100MHz, while this gpu rarely goes half a day if over 900MHz.

The gpu is the only one present in the system. It occupies a motherboard PCIe slot (no extender present in the system). The system has a 1100W power supply, so should have ample power for the gpu. GpuOwl V6.11-380-g36f4e2a is being run.

The system has ECC ram for reliability.
It is a Lenovo D30 with dual e5-2697v2 cpus. The OS is Windows 10 Pro x64, build 19041.867.

The preceding exponent 337333333 accumulated 144 GEC error detections during its primality test. While running a single long exponent (340000607), I have been slowly changing and logging changes to gpu ram clock rate, seeking an acceptably low error rate, over a period of weeks, and will continue until it reaches an acceptable error rate or gains cease. It has completed M340000607. M480003217 is running now.

Gpuowl instances v6.11-380 and v7.2-53 have been adjusted to -log 20000 to take advantage of a slight performance optimization given its error rate. That's revisited as error rate changes. Also -block 1000 for future exponents, to reduce GEC frequency a little (reducing overhead by ~0.3%) from the current -block 400.

The gpuowl log contains timing of error detections but not cause.
Analysis of gpuowl logs shows a subset, about 12%, of the error detections are with res64 value zero.
There appear to be multiple error modes, since there are both zero residues and more commonly seemingly random residues when errors are detected.
There is no apparent correlation between errors and time of day.
There is no initial clear indication of nonrandomness to the res64s that are nonzero.
Using https://www.gigacalculator.com/calcu...calculator.php to evaluate the extremes of frequency observed in the first 33 error-detected res64 hex digits, the large number of zeros is certainly nonrandom. All other counts are within the 95% confidence level of being random occurrence in the 528 digits.
Setting aside the 80 digits due to the 4 all-zero errors included in the digit frequency analysis as nonrandom, sample size declines to 464, and the 95% thresholds are 20 to 37 occurrences of a given digit value. The number of occurrences of the following digits fall outside the 95% probability of chance count range (20-37):
Code:
digit count
1     40
7     41
9     40
c     18
The difference between the total zero count and those from the 4 all-zero occurrences is only 20; the probability of only 20 or less is ~4.6%.
The total number of odds or of evens is only 3 away from equal (261 vs 267) which is suspicious given that 64 are nonrandom even zeroes. Subtracting them out there are 261 odds vs 203 evens in 464. A 0.4% probability at 50-50 chance.

Overall during the run, average time betweeen PRP error detections was ~8 hours 19 minutes (8.32 hours)

For a similar error rate, due to duration of run, LL at the DC wavefront without Jacobi check requiring ~7 hours would sometimes succeed and sometimes not. LL at the first test wavefront requiring ~28 hours would probably fail every time, without Jacobi check, and most times with it.(0.528/8.32 ~0.097 chance of successful completion with Jacobi check). A 100Mdigit run at ~15 days would have a very slim chance of completing correctly (0.515*24/8.32 ~9.4E-14)
Attached Files
 time of ee occurrence M340M.pdf (35.5 KB, 11 views) time of ee occurrence M340M histogram.pdf (992.1 KB, 13 views)

Last fiddled with by kriesel on 2021-03-31 at 12:46

 Similar Threads Thread Thread Starter Forum Replies Last Post kriesel Software 49 2018-08-15 23:28 ET_ FermatSearch 5 2016-07-03 10:58 R.D. Silverman GMP-ECM 11 2013-06-29 20:34 Maximus Math 15 2011-11-08 19:38 T.Rex Math 1 2010-01-03 11:34

All times are UTC. The time now is 03:45.

Sat May 8 03:45:51 UTC 2021 up 29 days, 22:26, 0 users, load averages: 1.19, 1.38, 1.41