GFN Prime Search Sieving

This sieve supports the GFN Prime Search located in PRPNet. Ports currently available for PRPing are:

port 12005: GFN32768
port 12003: GFN65536
port 11002: GFN262144
port 11001: GFN524288

Sieving is active for the following

N = 1048576
N = 2097152
N = 4194304

This sieve uses David Underbakke's AthGfn64 sieve program (currently available only for Windows 64 bit). It is GUI based and the process is straight forward. Reservations are open in the PST forum. You'll need to register for access.

If you are interested, check out the instructions. PSA credit is available.

The effort at N = 4194304 has the potential to produce the World's Largest Prime.
____________

Here's an idea of the resources needed for these sieves:

A 2600k host with no HT can complete 2.25P/day each on 4 cores. That's 9P/day for 1 host. Sieve depth goals for the following are:

N = 1048576; p=4612P
N = 2097152; p=6918P
N = 4194304; p=9223P

N = 1048576 needs 18 2600k hosts to complete in 30 days
or 9 in 60 days OR 5 in 90 days

N = 2097152 needs 26 2600k hosts to complete in 30 days
or 13 in 60 days OR 7 in 90 days

N = 4194304 needs 35 2600k hosts to complete in 30 days
or 18 in 60 days OR 9 in 90 days

So with a three month window, 21 hosts could complete ALL 3 sieves. Pushing it out to six months would only require 11 hosts.

Therefore, it's not unrealistic to think that we can complete these sieves in a relatively short time span. Please spread the word...the more the merrier...and, after-all, it is the holiday season!!! :)
____________

Size of potential primes

N=1048576 & 2097152 early in their searches will yield Top 20 primes. However, both quickly advance to Top 10 primes at b~5475 and b~75 respectively.

Currently, Mersenne primes dominate the Top 10 with 9 primes. A Sierpinski Problem prime comes in at #10. Therefore, a prime in the GFN search with bases above the thresholds mentioned earlier would certainly be an AMAZING accomplishment.

We have to go back to June 2005 to find a non-Mersenne, non-Sierpinski Problem prime in the Top 10. Daniel Heuer's Generalized Fermat prime (1372930^131072+1) was pushed out by Derek Gordon's Sierpinski Problem prime (27653*2^9167433+1).

It would be nice to get another Generalized Fermat prime in the Top 10. Back during the height of the previous GFN prime search effort (2003), there were 4 top ten primes. At that time, the list looked like this:

will join N = 4194304 for small range (initially)
____________
wbr, Me. Dead J. Dona

This project will be the best of the year ! Not with credits, since we had seen better times, but the best of the year with Fun factor.. cheers.

Indeed it should be. The currently running GFN searches have already turned up FOUR MEGA-PRIMES this year. When you consider how few people particpate in PRPNet as compared to BOINC, that's a remarkable accomplishment.

Indeed it should be. The currently running GFN searches have already turned up FOUR MEGA-PRIMES this year. When you consider how few people particpate in PRPNet as compared to BOINC, that's a remarkable accomplishment.

Should geneferCUDA ever get ported to BOINC, 2012 will be quite the exceptional year in regards to large primes. ;) The landscape of the Top 20 primes would change considerably.

Any thoughts on doing a primality test should we be lucky enough to find a 13 million digit PRP? Extrapolating the timing brings us to something in the vicinity of 4 months of crunching on a very fast CPU using PFGW64, which doesn't checkpoint. I could just see the cleaners unplugging the computer to run a vacuum cleaner 3.5 months into the run. :)

Does this work, that is provide correct output, under wine?
Just that my available hosts for the amount of time involved (5P on a single core takes about 3-4 days) are Linux.

Any thoughts on doing a primality test should we be lucky enough to find a 13 million digit PRP? Extrapolating the timing brings us to something in the vicinity of 4 months of crunching on a very fast CPU using PFGW64, which doesn't checkpoint. I could just see the cleaners unplugging the computer to run a vacuum cleaner 3.5 months into the run. :)

Obviously it would have to checkpoint...otherwise, it would be senseless to use pfgw. Even if it did checkpoint, 4 months is a substantial timeframe to prove primality. We'd definitely need to appeal to the larger prime community to see if there's other software/hardware we could use. But first, we need to find a PRP. :)

C:\prpclient-4.3.6-windows-gpu\prpclient-3-gpu>llr -d -q"67234^16384+1"
Base prime factor(s) taken : 33617
Starting N-1 prime test of 67234^16384+1
Using all-complex Core2 type-1 FFT length 16K, Pass1=64, Pass2=256, a = 3

67234^16384+1 may be prime, trying to compute gcd's
3^((N-1)/33617)-1 is coprime to N!

67234^16384+1 is prime! Time : 67.183 sec.

C:\prpclient-4.3.6-windows-gpu\prpclient-3-gpu>llr -d -q"67236^16384+1"
Base prime factor(s) taken : 431
Starting N-1 prime test of 67236^16384+1
Using all-complex Core2 type-1 FFT length 16K, Pass1=64, Pass2=256, a = 3


67236^16384+1 is not prime. RES64: D4A9A34F08AF910A. OLD64: 7DFCE9ED1A0EB31A
Time : 170.291 sec. (This time was obtained while the computer was already running at 100% CPU utilization, so it's much higher than it should be.)

C:\prpclient-4.3.6-windows-gpu\prpclient-3-gpu>llr -d -q"70906^32768+1"
Base prime factor(s) taken : 293
Starting N-1 prime test of 70906^32768+1
Using all-complex Core2 type-3 FFT length 32K, Pass1=128, Pass2=256, a = 3

70906^32768+1 may be prime, trying to compute gcd's
3^((N-1)/293)-1 is coprime to N!

70906^32768+1 is prime! Time : 294.510 sec.

Does anyone have knowledge of how well LLR performs with GFNs?

@John, OK. I want to help in this project. ( If I can )
Can I do something with an Lap top ? 1,6 Ghz ?
GenuineIntel / Intel(R) Core(TM) i7 CPU Q 720 @ 1.60GHz [Family 6 Model 30 Stepping 5]
(8 Prozessoren)

@John, OK. I want to help in this project.

Would be curious to know how fast it is under a virtual 64-bit windows but that would have to be a newer host.

Would be curious to know how fast it is under a virtual 64-bit windows but that would have to be a newer host.

All you need is a cpu with VT or IO-MMU and Virtualbox or VMware Player...

Would be curious to know how fast it is under a virtual 64-bit windows but that would have to be a newer host.

All you need is a cpu with VT or IO-MMU and Virtualbox or VMware Player...

I *think* he doesn't even need VT. I believe you only need VT if you want to emulate a 64 bit guest OS running under a 32 bit host OS.

Since he's running a 64 bit distro as the host, I believe you can run a 64 bit VM even without VT. (It's possible that I'm mistaken on this; it's been two or three years since I ran into this problem with some computers at work.) I know for a fact you can run a 32 bit guest OS under a 32 bit host OS without VT, but that's not exactly the same thing.

Would be curious to know how fast it is under a virtual 64-bit windows but that would have to be a newer host.

All you need is a cpu with VT or IO-MMU and Virtualbox or VMware Player...

I *think* he doesn't even need VT. I believe you only need VT if you want to emulate a 64 bit guest OS running under a 32 bit host OS.

Since he's running a 64 bit distro as the host, I believe you can run a 64 bit VM even without VT. (It's possible that I'm mistaken on this; it's been two or three years since I ran into this problem with some computers at work.) I know for a fact you can run a 32 bit guest OS under a 32 bit host OS without VT, but that's not exactly the same thing.

VMware released the KB-entry Hardware and firmware requirements for 64-bit guest operating systems.

Virtualbox also needs VT or equivalent for a 64-bit guest OS.
But will virtualization run the sieve faster than Wine?

Would be curious to know how fast it is under a virtual 64-bit windows but that would have to be a newer host.

All you need is a cpu with VT or IO-MMU and Virtualbox or VMware Player...

I *think* he doesn't even need VT. I believe you only need VT if you want to emulate a 64 bit guest OS running under a 32 bit host OS.
Since he's running a 64 bit distro as the host, I believe you can run a 64 bit VM even without VT. (It's possible that I'm mistaken on this; it's been two or three years since I ran into this problem with some computers at work.) I know for a fact you can run a 32 bit guest OS under a 32 bit host OS without VT, but that's not exactly the same thing.

VMware released the KB-entry Hardware and firmware requirements for 64-bit guest operating systems.

Virtualbox also needs VT or equivalent for a 64-bit guest OS.
But will virtualization run the sieve faster than Wine?

There shouldn't be any inherent speed difference. In both cases, the CPU is executing the code directly, and there are very few OS or API calls -- it's all algorithm. The only difference in speed should be the overhead of the operating system, and that's got more to do with what kind of stuff you have running in the background than the environment you're running it under.

There are numerous essentially handwaving arguments that Wine should be faster than an virtual machine since it does not have to implement a complete environment including the OS. But what counts are numbers. My speed under Wine seems far too low as for that host I would have expected closer to 2P per day rather than just over 1P per day.

If you want to have some computation times, look at my hosts. Only the host with a GPU is baremetal, all others were migrated (P2V) sometime in the past.

This sieve (AthGfn64) is offline and requires 64-bit Windows. I just wanted to avoid the considerable down time in setting up a virtual machine if there was no performance gain.

• Installing a guest via an ISO is much easier and faster than using the cd/dvd-drive inside the host.
  
  
• Installing singlecore-VMs gives a better responsiveness as using multicore-VMs. Same for RAM. Only as much as needed not as much as possible.
  
  
• If you awaiting high cpu loads, install no more VMs than available cores minus 1. The host-os needs also some computation time. No availability of free host cpu cycles causes ever also a slow down of all guests.
  
  
• W7/W2k8-R2 (a service pack doesn't matter) as host-os are known to have problems (guest instabilities or losing of guest-network), if you give summarized more than 40% of host RAM to your guests. In most cases should be 1GB more than enough. In dependence of the used workstation/player version you will hit another known problem, if you give a virtual machine more than that. The most part of the guest RAM will be marked as pageable and fills only the virtual memory of the host. This causes further performance impacts
  
  
• Don't use sparse/growing-disks or snapshots. Both cause a decrease in writing-speed inside a vm.
  

____________
Best wishes. Knowledge is power. by jjwhalen

I do know that will be down time because just rebooting the host to ensure VT is enabled is significant down time!

Maybe I misunderstood that KB, but didn't it say that you do NOT need VT unless you're trying to run Mac OS as the guest OS?

Since the whole point of this discussion is to get Windows running, isn't the question about having VT moot?
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Unfortunately that cannot check if VT is enabled in the bios.

Also installing virtual machines under Linux is not trivial - mainly time consuming.

Maybe I misunderstood that KB, but didn't it say that you do NOT need VT unless you're trying to run Mac OS as the guest OS?

Since the whole point of this discussion is to get Windows running, isn't the question about having VT moot?

Unfortunately that cannot check if VT is enabled in the bios.

False, it was released at the end of 2005 and does not know anything about the existence of newer processors since then. The only purpose of this tool is to check whether the cpu has VT-support enabled or not by reading the MSR-registers inside the cpu.

Just totally wrong mate - RTFM (emphasis added):

Note: On hosts with EM64T VT-capable processors, you may not be able to power on a 64-bit
guest, even though the processor check utility indicates that the processor is supported for 64-
bit guests. VT functionality can be disabled via the BIOS, but the processor check utility cannot
read the appropriate model-specific register (MSR) to detect that the VT functionality has been
disabled in the BIOS.

I'm afraid I don't understand your question. Are you asking about Sieving? If so, what does that have to do with a spreadsheet?
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

For GFN we're searching for primes of the form b^2^n+1. The b range they are sieving on the PST forum is given in the heading of all the factor files:

Fair play.

Just out of interest, what is the optimal sieve depth for GFN N=20? I know the current "goal" is 50E, but also realise that doesn't mean it's the optimal depth. I'm going to keep ploughing away at it for a while yet, though!

Why are they sieving N=20 to b=100M when the highest b limit, (for Genefer80), is 24,500,000?

They are sieving N=22 to b=100M as well, even through the max b limit, (for Genefer80), is 16,290,000?

While the question you asked was "When should we stop sieving?", another related question is "Which N should I sieve?"

That's answered not with "Optimal Depth" but with "Genefer Hours Saved per Sieving Hour" or just "Hour per Hour" (h/h).

The h/h (and optimal sieve depth) for the different n's are as follows:

n:20: 2.4 (79E)

n=21: 15 (528E)

n=22: 13 (1740E or 1.74Z)

From this perspective, n=21 is the most efficient N to be sieving right now.

(I've revised the optimal sieve depth from what's in the previous post.)
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Why are they sieving N=20 to b=100M when the highest b limit, (for Genefer80), is 24,500,000?

GFN n=18 Sieve Starting

We are opening manual GFN sieving for N=18 (n=262144).

Please see http://primesearchteam.com/showthread.php?t=80 for reserving a sieving range. There are links on that page that take you to instructions as well as downloading the necessary software.

The sieve software is ONLY available for Nvidia GPUs running under Windows.

If you're wondering why we're opening this up to sieving now, it's because n=18 was originally sieved with the older CPU sieve program. We now have a much faster GPU sieve which also is capable of sieving much deeper than the old sieve. We haven't previously sieved n=18 with the GPU sieve because n=18, until recently, was being crunched with GeneferCUDA. The CUDA program is much faster than the CPU program we have to use now, which means the optimal sieving point was lower and there was no need to sieve deeper. With the "really fast GPU sieve + really slow CPU Genefer" condition we have now, it's worthwhile sieving again.

The sieve doesn't work below n=18, so we can't do additional sieving on n=15 or n=16.

n=19 is already sieved to optimal depth for GeneferCUDA (and is sieved much deeper than n=18). When n=19 becomes a CPU-only project, we'll check to see if it makes sense to do additional sieving.

n=21 needs more sieving, but right now n=18 is the priority. N=22 also needs more sieving, but that's already sieved really deeply and we don't get many factors.
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

Is the source code available? If so, I could try to make a Linux build. Of course if Windows-specific APIs are used, I could be doomed to fail :-) I'll pass on the VM route.

Follow-up question then would be if there is a test suite (or whatever you want to call it) to verify that a new build is operating correctly. I see some factor lists posted elsewhere in this thread. Is matching those good enough?

--Gary

Is the source code available? If so, I could try to make a Linux build. Of course if Windows-specific APIs are used, I could be doomed to fail :-) I'll pass on the VM route.

Follow-up question then would be if there is a test suite (or whatever you want to call it) to verify that a new build is operating correctly. I see some factor lists posted elsewhere in this thread. Is matching those good enough?

--Gary

As far as I know, we don't have the source code.

My computer has 2 GPUs. When I run a single instance GFNSvCUDA on each GPU and I am not running anything on my CPU, each processes at roughly 32.6P/day. When I also run LLR on the CPU, the performance of GFNSvCUDA increases to roughly 38.0P/day. Does anyone know why this behavior is seen?
____________

My computer has 2 GPUs. When I run a single instance GFNSvCUDA on each GPU and I am not running anything on my CPU, each processes at roughly 32.6P/day. When I also run LLR on the CPU, the performance of GFNSvCUDA increases to roughly 38.0P/day. Does anyone know why this behavior is seen?

GFN Prime Search Sieving

This sieve supports the GFN Prime Search located in PRPNet. Ports currently available for PRPing are:

port 12005: GFN32768
port 12003: GFN65536
port 11002: GFN262144
port 11001: GFN524288

GFN Prime Search Sieving

This sieve supports the GFN Prime Search located in PRPNet. Ports currently available for PRPing are:

port 12005: GFN32768
port 12003: GFN65536
port 11002: GFN262144
port 11001: GFN524288

all these links do not work

Is it known that the links to stats on the Manual Sieving page don't work? I'm currently getting a 404 error when trying to visit this link.
____________

Is it known that the links to stats on the Manual Sieving page don't work? I'm currently getting a 404 error when trying to visit this link.

Is it known that the links to stats on the Manual Sieving page don't work? I'm currently getting a 404 error when trying to visit this link.

For me, all of the links to stats on the manual sieving page work -- The link you provided did not, but if you notice you are trying to go the https version of the page, which doesn't seem to exist. So go to the same page, but instead use http:// instead of https:// and it'll work.

Is it known that the links to stats on the Manual Sieving page don't work? I'm currently getting a 404 error when trying to visit this link.

Considering that the current range for n-22 being checked for primes is 54650 thru 63792, shouldn't stats of recently removed entries from 5000 to 10000, be changed to something relevant? Such as 63000 to 75000?

The "Instructions" link in the first post on this thread no longer works. Where should I go to find out more if I want to get involved?