1)
Message boards :
General discussion :
Sieving - is it a continuous process, or is it relating to a specific k?
(Message 54701)
Posted 4036 days ago by  [SG]Puzzle-Peter
These threads are impossible to make sense of with the OP and his opening questions gone. So please delete them altogether.
|
2)
Message boards :
General discussion :
Sieving - is it a continuous process, or is it relating to a specific k?
(Message 54568)
Posted 4038 days ago by [SG]Puzzle-Peter
finnenummer,
I am not sure I really understand what you are asking, so I'll keep my answer rather general hoping to cover your point.
A sieve task takes all (rather small) prime numbers P in a given range and tests if they divide the candidates in the sieve file. If a P is found so that
P | k * 2 ^ n - 1, the candidate (i.e. this specific combination of k and n value) is removed from the sieve. Every small prime P in the task range is tested for every k-n combination in the candidate list (i.e. sieve file)
After sieving to a point where it is not efficient any more, the remaining candidates for which no factor was found are primarily tested one by one. These tests are LLR tasks.
The main difference between TRP and PPS is that TRP only needs to find ONE prime for each of the remaining k values. Once a prime is found, this k will not be tested for higher n values. PPS on the other hand is searching for ALL primes in a given range for k AND n. That's why there are lots of k values in PPS but only a small number (56) in TRP. All k values that are worked on are present in the sieves, so right now there will be 56 k values in a TRP sieve and they are all sieved in one go.
Hope that helps
Peter
|
3)
Message boards :
Sieving :
Fastest way to sieve on CPU?
(Message 53955)
Posted 4048 days ago by [SG]Puzzle-Peter
Also, how do you make it run properly (sr1sieve)? I can't get it to work.
This may be trivial, but just to make sure I will add this to what Lennart already posted: make sure your input file has the correct format (.npg). If not, use srfile to convert.
I would also suggest switching to sr1sieve sooner than p=1e12 as it is faster than srsieve for a single k.
|
4)
Message boards :
General discussion :
An equivalent problem for decomposing large numbers
(Message 52938)
Posted 4073 days ago by [SG]Puzzle-Peter
I can't judge this, I can only propose to post this at mersenneforum.org. (See link on the left under "Community") They have a math section and also a factoring section. Several members are mathematicians.
|
5)
Message boards :
Project Staging Area :
AVX build of llr (20%-50% faster)
(Message 48729)
Posted 4140 days ago by [SG]Puzzle-Peter
AVX comes with Intels Sandy Bridge processors; maybe this list helps
Thanks! So no speedup is the expected behavior. Still it should work without strange messages, right?
|
6)
Message boards :
Project Staging Area :
AVX build of llr (20%-50% faster)
(Message 48719)
Posted 4140 days ago by [SG]Puzzle-Peter
I tried llravx on linux and get occasional "segmentation fault" when running tests on bases other than 2.
So far I always ran them in background mode. I'll try to get one in foreground (may take some patience) and provide more detailed error messages (if any occur) as well as detailed machine and OS specs.
Alright, it took quite a while to realise what happened. It may be due to my way of doing this. Here's what I do:
I run llr manually during the night hours. In the morning I send a kill signal to the jobs, causing them to write a restart file and shut down. In the evening I restart the jobs.
Sometimes when doing this I get something like this:
./llravx llr_sr366_220k-222k.txt
Caught signal. Terminating.
*** glibc detected *** ./llravx: double free or corruption (out): 0x0b763150 ***
======= Backtrace: =========
[0x9d27a2e]
[0x9d2bde9]
[0x809528a]
[0x80853e9]
[0x8086e3e]
[0x8087e37]
[0x808a9d1]
[0x808ad76]
[0x9d1780f]
[0x8048201]
======= Memory map: ========
08048000-09da5000 r-xp 00000000 00:21 1300688712
09da5000-09db9000 rwxp 01d5c000 00:21 1300688712
09db9000-09dc7000 rwxp 09db9000 00:00 0
0b6ae000-0b915000 rwxp 0b6ae000 00:00 0 [heap]
f2900000-f2921000 rwxp f2900000 00:00 0
f2921000-f2a00000 ---p f2921000 00:00 0
f2ac9000-f31a7000 rwxp f2ac9000 00:00 0
f32a8000-f4de6000 rwxp f32a8000 00:00 0
f4de6000-f4de7000 ---p f4de6000 00:00 0
f4de7000-f57e7000 rwxp f4de7000 00:00 0
f57e7000-f57e8000 ---p f57e7000 00:00 0
f57e8000-f61e8000 rwxp f57e8000 00:00 0
f628a000-f75af000 rwxp f628a000 00:00 0
f75e9000-f75ea000 rwxp f75e9000 00:00 0
f75ea000-f75eb000 ---p f75ea000 00:00 0
f75eb000-f7feb000 rwxp f75eb000 00:00 0
ffea5000-ffed6000 rwxp 7ffffffcd000 00:00 0 [stack]
ffffe000-fffff000 r-xp ffffe000 00:00 0
Abort
16780.092u 2.091s 4:40:03.40 99.8%
OS: Linux 2.6.18-194.17.4.el5 #1 SMP Mon Oct 25 15:50:53 EDT 2010 x86_64 x86_64 x86_64 GNU/Linux
Machine: model name : Intel(R) Xeon(R) CPU X5670 @ 2.93GHz
Candidate tested: 1747*366^220569-1
When I tried to restart the job, it seemed to work fine. The candidate was processed to the end with no error messages. I did not have the time to compare residuals yet, but I will do this just to make sure.
This is a brand new and high-end dual-CPU machine but I don't see any speedup. So I suppose it does not support avx. I tried to find a list of avx-capable CPUs with no success. Could somebody tell me how to find out whether a machine does support avx?
Thanks
Peter
PS: interrupting llr jobs and restarting them with llravx seems to work just fine.
|
7)
Message boards :
Project Staging Area :
AVX build of llr (20%-50% faster)
(Message 47869)
Posted 4149 days ago by [SG]Puzzle-Peter
I tried llravx on linux and get occasional "segmentation fault" when running tests on bases other than 2.
So far I always ran them in background mode. I'll try to get one in foreground (may take some patience) and provide more detailed error messages (if any occur) as well as detailed machine and OS specs.
|
8)
Message boards :
Number crunching :
The Autumnal Equinox Challenge
(Message 41088)
Posted 4268 days ago by [SG]Puzzle-Peter
"I was getting dressed late one night
Along came a spider and I got afraid
I couldn't squash it flat, no matter how I tried
And when it looked me in the eye I ran away to hide..."
Not my words, but a rather funny song performed by the Toy Dolls. When I read "along come spiders" the memory popped up. Gotta look for the CD...
|
9)
Message boards :
Number crunching :
The Autumnal Equinox Challenge
(Message 40598)
Posted 4275 days ago by [SG]Puzzle-Peter
With WUs as long as Woodalls, some may consider even more "babysitting."
Imagine you have 90h left and your computer takes 60h per WU. You can crunch 1.5 WUs per core in the time left, but that won't help.
But with a little manual intervention, there is a way to get more.
On a dual core you could finish 2 WUs and bring 2 more to 50% each. You get challenge points for 2 WUs. Now, 50% x 2 = 100%. Let's make use of that.
First run 2 WUs to 50% (we are at 30h now)
Suspend one of them. The core crunching it will start a new one.
After another 30h, this new WU is at 50%, the one you didn't suspend is finished. (total of 60h now)
Now, UNsuspend the one that has been interrupted at 50%
Now both cores have WUs that are at 50% (which will finish in time) AND you have another one that is already finished.
End result: Not 2x100% plus 2x50%, but 3x100%
NOTE: This is not my idea, but I forgot who mentioned it first.
|
10)
Message boards :
Number crunching :
The Dog Days of Summer Challenge
(Message 39967)
Posted 4294 days ago by [SG]Puzzle-Peter
Something for the developers to consider: Is it possible to detect workunits handled through proxies and have a class that counts only workunits that didn't go through proxies?
In my case it is not even a proxy. It is just a PC with BOINC installed. I let it download a huge number of WUs, put them on a USB stick, take them to other machines ehere they are processed and after they're finished I transfer them back. Even the BOINC on this single machine "thinks" it crunched them by itself. I don't think there's any way of telling what happened between download and upload.
|
