mersenneforum.org  

Go Back   mersenneforum.org > New To GIMPS? Start Here! > Information & Answers

Reply
 
Thread Tools
Old 2021-06-25, 08:57   #12
kruoli
 
kruoli's Avatar
 
"Oliver"
Sep 2017
Porta Westfalica, DE

2×3×127 Posts
Default

P-1 is still efficient on CPUs if you have some GB(s) of RAM to spare. You'll have to configure that in the GUI once, otherwise the program might only do stage 1 and that would be inefficient.

P-1 can be really rewarding in the long run.

Last fiddled with by kruoli on 2021-06-25 at 08:58 Reason: Clarifications.
kruoli is offline   Reply With Quote
Old 2021-06-25, 17:02   #13
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

22·7·211 Posts
Default

Quote:
Originally Posted by LaurV View Post
Please DON'T do TF on a CPU. As other people told you, it will be a waste, it takes ages; a GPU is hundreds of times faster at this kind of job (unless you have a "xeon phi" or something).
I'm of the opinion that for TF work feasible by a GPU application (mfaktx or mmff), TF is also a waste of a Xeon Phi, which could be much more productive doing LL DC, PRP DC, P-1, first time PRP/GEC/proof, PRP/GEC/proof as DC of previous LL, or a blend of those using multiple workers. A no-DIMMs Xeon Phi can serve well with a P-1 worker using most of the MCDRAM, and the other workers running DP-intensive LL DC or PRP with modest memory requirements. A Xeon Phi 7250 is about the equal of half a Radeon VII for LLDC, PRP or P-1. TF is best left to NVIDIA GPUs, particularly GTX 16xx, RTX20xx or RTX30xx with their very low DP performance and high TF performance.

On occasion for special tasks, TF for over-32-bit exponents can be run (mfactor, factor5) alongside DP-intensive mprime / prime95 or mlucas work, on CPUs with hyperthreading. PRP or LL is quite memory intensive, TF is not. Doing this impacts prime95 throughput somewhat.

TF can be run on the IGP in a cpu package via mfakto, on some IGP models. That also impacts prime95 throughput, through the total package power budget. IGP throughput in many cases is 10-200x slower than a discrete GPU's performance. It can be useful as a way to get familiar with GPU TF computing before buying a compute-capable GPU or while waiting for delivery. Test whether it is worthwhile.

Last fiddled with by kriesel on 2021-06-25 at 17:04
kriesel is offline   Reply With Quote
Old 2021-06-25, 18:01   #14
birtwistlecaleb
 
birtwistlecaleb's Avatar
 
Jun 2021

23·7 Posts
Default

Quote:
Originally Posted by kruoli
P-1 is still efficient on CPUs if you have some GB(s) of RAM to spare.
Is 6 ok for temporary disk space?
birtwistlecaleb is offline   Reply With Quote
Old 2021-06-25, 18:04   #15
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

277016 Posts
Default

P-1 uses RAM (main memory) not so much disk space. The more RAM you have the more useful P-1 work you can do more efficiently.

Of late the generation of proof files while doing PRP requires more disk space.
Uncwilly is offline   Reply With Quote
Old 2021-06-25, 19:56   #16
birtwistlecaleb
 
birtwistlecaleb's Avatar
 
Jun 2021

5610 Posts
Default

Thanks for answering! I am ok if this gets locked now.
birtwistlecaleb is offline   Reply With Quote
Old 2021-06-25, 20:40   #17
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

22×7×211 Posts
Default

Quote:
Originally Posted by birtwistlecaleb View Post
Is 6 ok for temporary disk space?
Depends on what you're attempting.
P-1 up to 1G in prime95 with 3 backup files total occupies less than 1GB of disk space per exponent.
PRP with proof power 8 for 100M exponent takes ~3.2GB for the proof residues file; power 9, 6.4GB. That is proportional to exponent and exponential with proof power (c p 2power) . So power 10 on 1G would require ~130. GB.
https://www.mersenneforum.org/showpo...45&postcount=4
Please use the reference info for such questions.
kriesel is offline   Reply With Quote
Old 2021-06-26, 07:15   #18
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
"name field"
Jun 2011
Thailand

264E16 Posts
Default

Quote:
Originally Posted by birtwistlecaleb View Post
Is 6 ok for temporary disk space?
My understanding is that that 6 is there for the case "things go bad". It is not related to P-1 stage 2, or ECM. Let it 6.
LaurV is offline   Reply With Quote
Old 2021-06-26, 15:29   #19
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

171416 Posts
Default

Quote:
Originally Posted by LaurV View Post
My understanding is that that 6 is there for the case "things go bad".
No, that case of things going bad would be "max emergency memory" setting. You seem to be conflating multiple settings. Quoting from the prime95 readme.txt:
Code:
CONTROLLING RESOURCE USAGE
--------------------------

The Options/Resource Limits dialog box contains important settings to control
the program's resource usage.

The PRP primality test generates a proof file that eliminates the need to double-check your results.
Generating these proofs requires lots of temporary disk space.  Proof files must be uploaded to the
server which could be an issue for some users with limited bandwidth.

PRP proofs are a little more efficient when using a larger proof power.  However, this requires more
disk space and internet bandwidth.  The table below shows resource usage for several proof powers
and exponents.

Proof                              Temp disk space / Proof file size
Power         Exponent=100,000,000         Exponent=200,000,000            Exponent=332,000,000
7               1.6GB / 200MB*                 3.2GB / 400MB*                 5.3GB / 664MB*
8               3.2GB / 113MB                  6.4GB / 225MB                  10.6GB / 373MB
9               6.4GB / 125MB                  12.8GB / 250MB                 21.2GB / 415MB
10             12.8GB / 138MB                  25.6GB / 275MB                 42.4GB / 457MB

After the proof is uploaded to the server it is pre-processed creating a smaller file that
must be downloaded by users doing proof certification work.  Proof certification work is
very fast (256, 512, or 1024 times faster than the original PRP test depending on proof
power).  By default, your computer is signed up for occasional proof certification work.

*For proof power 7, it is possible to cut the proof file size in half but this doubles the
work required by the user doing the certification.  See undoc.txt.

Temporary disk usage
--------------------

This setting limits the amount of temporary disk space used BY EACH WORKER doing PRP work.
See the table above for how that affects proof power.  A proof power of 8 or more is desirable,
but a proof power of 7 is acceptable.  The default limit of 6GB will use a proof power of 8
for exponents you are likely to be assigned for first time tests over the next several years.
This can be left at 6GB for PRP up to 100Mdigit (332M+ exponent), although you'll be doing the eventual Cert runner a favor by using higher proof power than 7 if you have 11 or 22 GB of HD to spare in the 100MDigit case, or allocating 7-8GB for ordinary production wavefront PRP&proof to support power 9. Many of us will have enough space in these days of TB boot drives. Just like it says, this relates to producing PRP proofs. Not P-1 factoring, not LL tests, ...


Code:
Daytime and nighttime P-1/ECM stage 2 memory
--------------------------------------------

On occasion, you may be assigned an exponent that needs P-1 factoring prior to running a
primality test.  Or you might have chosen to do P-1 or ECM work (possibly because you do
not want to devote disk space for large PRP files).  In these situations, the program can
do stage 2 slightly more effectively if it is given more memory to work with.  However,
if you let the program use too much memory then the performance of ALL programs will suffer
due to thrashing.  

That is, most of the time this setting is not used and even with minimal settings the program
will work just fine.  Should you decide to change these settings, how do you choose intelligently?
Below are some steps you might take to figure this out:

1)  Be conservative.  It is better to set the memory too low than too high.  Setting the value
too high can cause thrashing which slows down all programs.

2)  Start with how much memory is installed in your machine.  Allow a reasonable amount of memory
for the OS and whatever background tasks you run (say 0.5 to 2.0GB).  This represents the maximum
value you should use.  The program won't let you enter more than 90% of installed memory.

3)  Assuming you run your machine 24 hours a day, what hours of the day do you not use your computer?
Make these your nighttime hours and let the program use a lot of memory during these hours.  But
reduce this value if you also run batch jobs at night.

4)  Factor in the information below about minimum, reasonable, and desirable memory amounts for some
sample exponents.  If you choose a value below the minimum, that is OK.  The program will simply skip
stage 2 of P-1 factoring.

    Exponent    Minimum        Reasonable    Desirable
    --------    -------        ----------    ---------
    100000000     0.2GB          0.7GB         1.1GB
    333000000     0.7GB          2.1GB         3.5GB

For example, my machine is a dual-processor with 8GB of memory.  I guess Windows and the programs I
normally use can survive on 2GB of memory.  Thus, I set memory to 6.0GB.  This is my nighttime setting.
During the day, I run more programs, so I set memory to 1.5GB.  I can always stop prime95 if it is
doing stage 2 P-1 factoring and I suspect memory is thrashing.  More casual users might want to set the
daytime memory to 0.5GB so they never have to worry about prime95 impacting system performance.
For long running P-1 assignments, set daytime and nighttime to the same memory amount. Stage 2 may restart from the beginning when the memory allowance changes. If you have the ram to spare, experiment with lots of it. On a 16GB ram system, I have it set to 12 and 12. (But I'm doing P-1 on exponents 104M - 999.99M.) With lots allowed, prime95 seems to do a good job of taking what's useful and available and leaving the rest. Default 0.3GB is usually too small. It was chosen small so novices don't get a bad initial impression of prime95 from a large default possibly causing thrashing on a small-ram system.


Code:
Max emergency memory
--------------------

If the program cannot write to the large temporary file, it will use emergency memory to hold data in
hopes that it can later be successfully written to the temporary file.  For me, this gives me about a day
to correct a network drive that has gone offline.  If you use a local disk to store your large temporary
proof files then emergency memory might be used if the local disk is full.
This is ram, not disk, for when there's a disk issue, to cache in ram what could not be written to disk, in the hope of being able to write it to disk later. Nice feature George! Because of the reference to "large temporary file" or "large temporary proof files", this appears to apply to PRP proof runs only. Default is 1GB I think.
Attached Thumbnails
Click image for larger version

Name:	temp disk for prp and ram for p-1 s2 or ecm.png
Views:	42
Size:	20.7 KB
ID:	25190   Click image for larger version

Name:	max emergency memory etc.png
Views:	45
Size:	20.6 KB
ID:	25191  

Last fiddled with by kriesel on 2021-06-26 at 15:59 Reason: add pics
kriesel is offline   Reply With Quote
Old 2021-06-27, 00:12   #20
birtwistlecaleb
 
birtwistlecaleb's Avatar
 
Jun 2021

708 Posts
Default

Thanks!
birtwistlecaleb is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Manual GPU assignment page down? Ensigm PrimeNet 3 2020-10-13 13:02
Manual assignment bug? greg PrimeNet 3 2020-05-07 16:58
Whoops! Forgot to extend manual LL exponent... Rodrigo PrimeNet 26 2013-01-01 13:35
Manual assignment Unregistered Information & Answers 12 2010-10-07 16:15
Problem with LL manual assignment JuanTutors Software 2 2008-12-02 01:57

All times are UTC. The time now is 09:13.


Sat Nov 27 09:13:29 UTC 2021 up 127 days, 3:42, 0 users, load averages: 1.33, 1.24, 1.21

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.