The Scottish Boinc Team

Hi all,

Need someone in UK who has at least a machine with 32 cores or more (AVX2.0), more than 32GB of memory, to run some work from a project I am involved with, I'll pay for the electricity.

Please PM so I can send you the details.

Carlos

PS(Binaries I have are windows for ivybridge, sandybridge and haswell Intel processors)

Sorry Carlos, I have no such systems. Best I can offer is Xeon x5650 with 24 threads and it is linux .

I don't think we've got anyone in the UK with that many cores in a single machine Carlos

Thank you guys.

Does the system need to be in the UK?

scole250 wrote: ↑Tue Jun 13, 2017 8:04 pm Does the system need to be in the UK?

It's easier to pay for the electricity by bank transfer.

How much system time do you need? And how soon you need it?

scole250 wrote: ↑Tue Jun 13, 2017 8:36 pm How much system time do you need?

Running on my laptop 4 cores (Ivybridge) is around 64 days. Binaries attached,dll should be on the exe folder.

To run use the current binary, windows create start.bat as:

start /low /min gap5_f_ivybridge.exe -n1 4800e15 -n2 4900e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4

use the correct exe file
mem - as max as possible, this case I am using 12GB out of 16GB
-t - number of threads, using 4
speed I get is around 18E9n/sec

(100E15/18E9)/(24*3600)=64 days

The 100e15 range can be broken:

start /low /min gap5_f_ivybridge.exe -n1 4800e15 -n2 4810e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4810e15 -n2 4820e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4820e15 -n2 4830e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4830e15 -n2 4840e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4840e15 -n2 4850e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4850e15 -n2 4860e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4860e15 -n2 4870e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4870e15 -n2 4880e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4880e15 -n2 4890e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4
start /low /min gap5_f_ivybridge.exe -n1 4890e15 -n2 4900e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4

Does it compute faster if you turn hyper threading off?

Bryan wrote: ↑Tue Jun 13, 2017 10:09 pm Does it compute faster if you turn hyper threading off?

It's faster with HT on.

Can you break it into smaller chucks? ie if you change your .bat file to


start /low /min gap5_f_ivybridge.exe -n1 4800e15 -n2 4850e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4

would it just run the 1st 5 and then stop?

Does it upload the results automatically or what would someone do with them?

They are standalone clients without communication with any server.
Yes, you can breakdown the range into smaller ones as I have on my post. Just change the number after -n1 and after -n2 respectively. Your example does half of the range, don't forget to change -t accordingly and -mem as well.

Solutions are saved on gap_solutions.txt and check points every 45 mins (it's in function of the size, can't remember, that's the time on my laptop), file worktodo_gap.txt.

start /low /min gap5_f_ivybridge.exe -n1 4800e15 -n2 4850e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 4

This one does one junk on 50 and then it closes running on 4 threads and using 12GB of memory. It starts the range from 4800 to 4850, but you can change from 4800 to 4810, 4810 to 4820, and so on. Each block of 10 takes 6 days on my ivibridge machine using 4 threads. Haswell machines are at least 50-70% faster than my laptop.

Okay Carlos, I have 64 threads and 128G. What do I need for a bat file to run the 1st one ... just the 1st segment. I want to see how long it takes before I commit beyond that 1 slice.

I'm guessing:

start /low /min gap5_f_ivybridge.exe -n1 4800e15 -n2 4810e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 128 -t 64

Bryan, that's correct just decrease memory a little to 120.
Also put the dll inside the folder and bear in mind the binary to use in function of the CPU you have.

Okay I'll give it a try. That machine is on Linux right now so it will take me a little bit to get going. Is there any way to tell the progress or does it just stop?

On the command line you will see the progress on the number starting on 4800e15. There is no % progress. Also it will show you the speed in xxe9 n/sec.
You also can open the solution txt file to see the results getting closer to the n2 limit. Think you are going to do that block in less than one day.
The client will close down when it finishes the range.

What's the processor type and speed?

E5 2683 V4 @ 2.6G with 64 threads under load.

Think you will do that range of 10 in less than 12 hours. Fingers crossed. Use the haswell binary.

I'd set the memory to something less than 128GB. 120GB max, maybe even 112GB

I set the memory at 120G and the 1st segment is now cranking. At the 5 minute mark the CPU is at 65-80% with 56G in use.

The 1st checkpoint came up and it appears it is less than double the speed of your laptop

32.67e9 vs your 18e9

The machine at 7505 seconds was at 4800245xxx so if the end result is 4810 and it is linear then it will take 85 hours to finish 4800-4810.

I see that the 2 CPUs are not being utilized equally. If I were to do it again then I would break the 4800-4810 into 2 WU and run 1 on each CPU. To do that you might need to launch the 2nd one from a different directory so it doesn't get screwed up with writing to the results file. It is now using 85G of memory.

By the way, it is using AVX2 because the CPUs are downclocked by 250MHz.

If my systems are up to helping , then let me know and i'm up for helping with no payment required.

Sorry for coming back so late. Please find the linux binaries.
I did inquire the people running dual socket and they are running two instances of the client in separate folders.

Carlos, my systems are NOT using AVX. I tried 2 more systems E5 2695 V4 (72t) and E5 2695 V3 (56t) and according to the clock rate they aren't using AVX with the Haswell executable.

To confirm I ran BOINC Stop on both machines and looked at the clock rate ... ie on the V3 machine it was 2.8GHz. I then switched to Asteroids which is AVX and the clock rate dropped to 2.5GHz as it should.

When I start the Gap application the clock rate is the same as standard ... they don't downclock which means they are not running AVX/FMA3.

It is possible that the application is doing a lot of preliminary work and then doing the actual AVX calculation in bursts. I did notice that it runs for about 2 hours at 80% CPU usage and then does a 5 minute burst where it goes to 100%. Right after than it does the status update.

If that 64 thread machine is going to produce less results than 2 of your 4 core laptops I'm not inclined to continue.

Bryan, got confirmation from Thomas that: "Actually the AVX registers are only used once for the translation of the floating point input arguments (n1, n2) into integers." which follows the behavior you're experience. AVX is not being fully utilized. Apologies and please fell free to stop the running.

Okay Carlos and thanks for checking. I will be shutting it down since I don't see much benefit of running a machine of that class.

Anyone here with a fast intel quad (i7) or Ryzen (8C/16T) machine that could run me the following:

start /low /min gap5_f_XXX.exe -n1 4500e15 -n2 4510e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 8

start /low /min gap5_f_XXX.exe -n1 4500e15 -n2 4510e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 12 -t 16

Here's the link for the 'First occurrence prime gaps' where it can be seen the works carried out since 1878.
http://www.trnicely.net/gaps/gaplist.html

Currently mersenneforum is working from 4000e15 to upper bound 5000e15 and the client can be used up to 16950e18. The code will have to be rewritten to be able to run beyond 16950e18.
This type of work is the kind you will be remember for decades...Please let me know if TSBT and SETI.USA are willing to support this project, there is no rewards, no BOINC points, no automate reservation system, it's all manual.

Thank you in advance,

Carlos

Bryan wrote: ↑Wed Jun 14, 2017 8:08 pm Okay Carlos and thanks for checking. I will be shutting it down since I don't see much benefit of running a machine of that class.

Does it use all the threads and memory?

It uses all threads but it's not well optimised therefore you better run a client per 4 threads. You can play along with the client.

You can start with one folder running for example (16 thread machine, 32 GB):

start /low /min gap5_XXX.exe -n1 4500e15 -n2 4510e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 28 -t 32

Or run several folders each one with the following command:

start /low /min gap5_XXX.exe -n1 4500e15 -n2 4503e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 7 -t 4
start /low /min gap5_XXX.exe -n1 4503e15 -n2 4506e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 7 -t 4
start /low /min gap5_XXX.exe -n1 4506e15 -n2 4508e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 7 -t 4
start /low /min gap5_XXX.exe -n1 4508e15 -n2 4510e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 7 -t 4

Running 4 processes this way, each in it's own folder...

start /NODE 0 /AFFINITY 0xFFFFFFFFF /low /min gap5_f_haswell.exe -n1 4500e15 -n2 4503e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 28 -t 12
start /NODE 0 /AFFINITY 0xFFFFFFFFF /low /min gap5_f_haswell.exe -n1 4503e15 -n2 4506e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 28 -t 12
start /NODE 1 /AFFINITY 0xFFFFFFFFF /low /min gap5_f_haswell.exe -n1 4506e15 -n2 4508e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 28 -t 12
start /NODE 1 /AFFINITY 0xFFFFFFFFF /low /min gap5_f_haswell.exe -n1 4508e15 -n2 4510e15 -gap 1250 -delta 155 -sb 24 -bs 16 -mem 28 -t 12

Which speeds are you getting? What name should I add to the reservation list?
When all are finished can you please send me the gap_solutions.txt files to pinhodecarlos at yahoo dot com.

It's Xeon V4 (Broadwell EP) running at 2.8Ghz. Each of the 4 processes is getting a little better than 25e9 n/sec. Check my math but the 2 processes running the 3e15 ranges should take (3e15/25e9)/(24*3600), about 1.39 days, correct?

scole250 wrote: ↑Sun Jun 25, 2017 8:59 pm It's Xeon V4 (Broadwell EP) running at 2.8Ghz. Each of the 4 processes is getting a little better than 25e9 n/sec. Check my math but the 2 processes running the 3e15 ranges should take (3e15/25e9)/(24*3600), about 1.39 days, correct?

Calculations correct.

New windows binaries showing progress in hours. bs flag can be increases from 16 to 18, little speed increase.

Complete and results sent. Will I be famous?

scole250 wrote: ↑Tue Jun 27, 2017 11:18 am Complete and results sent. Will I be famous?

If you do what I've proposed by email, yes.
The link has been update but its missing the list of researchers. I've sent an update and will chase after my operation.

Here are some Linux binaries (Nehalem and Haswell) for Antonio's latest version gap5_g.

New flags to be used:

gap5_g -n1 XXXX -n2 XXXX -gap 1250 -delta 155 -sb 24 -bs 18 -t 4 -mem 12

What are the different linux files for and which is run on which system? What's the difference between the static and non-static files?

EDIT: I figured out what the static executable was for.

Here's the email from Dr. Thomas R. Nicely. We are making history!!!!

"This is in reply to your messages of 12 July and 17 July.

Your own new first known occurrence prime gap of 1406, and Steve Cole's
new first known occurrence prime gap of 1396, have now been posted at
[ http://www.trnicely.net/gaps/gaplist.html ]. The corresponding discoverer
abbreviations are Pinho_MF and ColeStev.

Thank you both for your contributions. I look forward to receiving additional
discoveries from you.

Regards,

TRN"

There's a client version which requires -mem (1.55 + 0.05t) to run, where t is the number of threads and it is less quick. (REQUIRES LESS MEMORY TO RUN)
Range 7000-8000 available.
start /low /min gap5_size.exe -n1 7000e15 -n2 8000e15 -gap 1346 -delta 155 -sb 24 -bs 18 -mem 4 -t 8

Just to share something interesting about Dr. Thomas R. Nicely. Original e-mail message announcing the discovery of the Pentium division flaw, 30 October 1994: http://www.trnicely.net/pentbug/bugmail1.html

Faster windows client, still awaiting for the linux one.

Are the command line arguments for gap5_h.exe and gap5_h_haswell.exe the same as the arguments for the g version?

scole250 wrote: ↑Tue Jul 25, 2017 9:22 pm Are the command line arguments for gap5_h.exe and gap5_h_haswell.exe the same as the arguments for the g version?

Yes.

New gap10 is 50% faster than gap5 since it now uses avx and a new algorithm. Went from 22e9 n/s to 31e9 n/s. Available ranges are:
7000-8000
9000-10000
10000-16000

Records found so far. Project keeps going on.