Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
Message boards :
Number crunching :
Beta testing of PRST app is open on "GFN Server"
| Author |
Message |
|---|
streamVolunteer moderator
Project administrator
Volunteer developer
Volunteer tester
Send message
Joined: 1 Mar 14
Posts: 1022
ID: 301928
Credit: 543,195,386
RAC: 1
                        
|
|
I've just created a new project, "PRST testing" to test new prime searching program - PRST.
PRST is a new prime testing program, written from scratch by Pavel Atnashev, author of LLR2. This is still a work in progress, but it's quite stable and basic types of tests are already implemented. The thing most important to us is that PRST can test factorial and primorial primes with fast double-check and Gerbicz-Li error detection.
As an essential step to move our factorial and primorial searches to Boinc, we must test new program and all brand new server scripts in real-life conditions. On the first phase, PrimeGrid PrpNet factorial and primorial work will be double-checked.
The testing will be run on my own "Private GFN Server". It's not on PrimeGrid yet. If you're already registered on my server, just go to preferences and select new application.
If you're not registered yet, you must go to the web page of my server http://boincvm.proxyma.ru:30080/test4vm/ and create an account there. The inviation code is PrimeGrid. Automatic registration from Boinc client probably will not work because, as far as I remember, it was not possible to specify invitation code during built-in registration. After registration, you can attach your Boinc client to new project using URL above.
The program is similar to LLR(2) and is CPU-only. Applications are available for Windows64 and Linux64 platforms.
Multithreading is recommended (currently 3 or 4 cores per task for most desktop CPUs, depending on CPU total number of cores) and supported using app_config.xml. See this post for details. | |
|
|
|
|
Just downloaded a batch of tests single core estimate 23 minutes on a 5900x. Does this need a variety of hardware? Should I switch up computers or is the first batch just DC of PSA projects not matter as much? Also if we do get a mismatched factorial DC will you be the 3rd party or does it create a 3rd wu? | |
|
|
|
|
Seeing all failed tasks ... across my various Windows hosts
Some with error 255 (0xff) Unknown error number
Others with error -1073741795 (0xffffffffc000001d) Unknown error number
ALL crash/fail within a few seconds of starting up.
____________
 | |
|
|
|
Just downloaded a batch of tests single core estimate 23 minutes on a 5900x. Does this need a variety of hardware? Should I switch up computers or is the first batch just DC of PSA projects not matter as much? Also if we do get a mismatched factorial DC will you be the 3rd party or does it create a 3rd wu?
The program uses the same GWnum library as LLR2, so the performance is similar. But we'd appreciate testing on unreliable/overheating hardware, frequent starts/stops to see how PRST handles situations and faults.
All tasks for the foreseeable future are DCs of the old factorial/primorial work. As with all DCing on Private GFN Server, in case of residue mismatch the fast-DC system is activated and the new residue is proven by the certificate. | |
|
streamVolunteer moderator
Project administrator
Volunteer developer
Volunteer tester
Send message
Joined: 1 Mar 14
Posts: 1022
ID: 301928
Credit: 543,195,386
RAC: 1
|
Just downloaded a batch of tests single core estimate 23 minutes on a 5900x. Does this need a variety of hardware? Should I switch up computers or is the first batch just DC of PSA projects not matter as much?
You can test in any style according to you personal preferences / goals. We need to check how application behaves in real life on a variety of different computers, where users are performing usual daily tasks (pausing / aborting applications, rebooting computer, installing Windows updates and so on), and how server handles data returned by application in these cases. Just run it as long as you want to, and report to forum or Discord if you've noticed something unusual.
Also if we do get a mismatched factorial DC will you be the 3rd party or does it create a 3rd wu?
"fast double check" system will be activated to prove that our task was correct and PrpNet task was wrong. It will create new short "certificate" tasks and send it to somebody. There are some threads/posts on forum which explains how fast DC works in LLR2 (PRST uses same algorithm), so I'll not repeat it here.
| |
|
|
|
|
Also, PRST is a native BOINC application, no wrapper is used. It's a new feature, needs to be tested thoroughly. | |
|
streamVolunteer moderator
Project administrator
Volunteer developer
Volunteer tester
Send message
Joined: 1 Mar 14
Posts: 1022
ID: 301928
Credit: 543,195,386
RAC: 1
|
Seeing all failed tasks ... across my various Windows hosts
Some with error 255 (0xff) Unknown error number
Others with error -1073741795 (0xffffffffc000001d) Unknown error number
ALL crash/fail within a few seconds of starting up.
I've installed new Windows build, hopefully it'll fix the program.
Look like previous version was erroneously built with AVX instructions, which are not supported by your CPU.
| |
|
|
|
|
Hello!
Thank You Stream and Pavel for making a new project available!
I've been running the test units since they appeared yesterday evening, about 100, mostly version 1.01 without any issues.
They all "survived" latest windows update that was installed during the night.
After restarts this morning, first closing Boinc nicely, still no errors or invalids occurred!
One of my pcs (id 1150) got special "horror" treatment!
First installing windows update and then restarting without closing Boinc and no stopping of running work units before shutting down, Boinc and all running units survived!
And last test was updating Intel Graphics with Boinc running. After install finished the on/off switch was used to restart everything, no errors !
If You ask me this seems rock safe!
Then a few questions:
How many units to Double Check?
You seem to start with factorial primes easier to test, primorial tests to follow?
I guess it is not DC until TdP starts in February?
They are both megaprimes if we start from where PRPNet are today?
Thanks again and regards,
Hans Sveen
Oslo
____________
MyStats
My Badges | |
|
 Michael Goetz Volunteer moderator
Project administrator
Send message
Joined: 21 Jan 10
Posts: 13955
ID: 53948
Credit: 392,890,430
RAC: 184,027
                         
|
Also, PRST is a native BOINC application, no wrapper is used. It's a new feature, needs to be tested thoroughly.
Nice!
____________
My lucky number is 75898524288+1 | |
|
streamVolunteer moderator
Project administrator
Volunteer developer
Volunteer tester
Send message
Joined: 1 Mar 14
Posts: 1022
ID: 301928
Credit: 543,195,386
RAC: 1
|
Then a few questions:
How many units to Double Check?
There are 176500 factorial tests from N=50K to N=300K (current PrpNet leading edge). Everything below 50M was tested manually. It's quite significant amount of work because factorials are growing very fast, both FFT size and runtime will be greatly increased soon. We jumped from 50M to 100M in 24 hours or so, but it'll slow down soon.
You seem to start with factorial primes easier to test, primorial tests to follow?
I just didn't preprocessed primorials yet. As soon as data will be ready, I'll load primorials as well. Probably task will be sorted by FFT size, so both types of primes will be tested at same time and tasks will have similar runtime.
I guess it is not DC until TdP starts in February?
I'm sorry but I didn't understood the question fully. Factorials and primorials (if any missing primes will be found) will not acceptable for 2023 TdP because this test project is running on other server. The goal is to run full-scale Boinc search on PrimeGrid, but it definitely will not happen before February.
They are both megaprimes if we start from where PRPNet are today?
Yes, both projects will be mega-primes.
| |
|
HonzaVolunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 15 Aug 05
Posts: 1949
ID: 352
Credit: 6,014,223,837
RAC: 1,553,618
                                
|
|
Great news!
I was trying to break it from command line with leading edge candidate.
Run a candidate for a short while, Ctrl+C.
Restarted: 103145!-1 restarting at 3.1%.
That's before 1/8 proof stage, let's say good checkpointing.
Run up to 22.7% stage, Ctrl+C.
Now it restarted from 12.5%, ie. first proof stage.
It is expected behaviour?
What is the best way to test MT?
I already know it doesn't scale great (-t 4 or -t 16) from command line.
AFAIK, there is no way to set web preferences for that.
app_config.xml?
Plan class?
Any template for that?
____________
My stats | |
|
|
|
|
Hi !
Thanks for aswering my questions!
Sorry for not proofreading well, yes you did understand mine qustions right,
I guess it is not DC until TdP starts in February? should have been written:
I guess the DC is not finished until TdP starts in February?
Hans S.
____________
MyStats
My Badges | |
|
|
|
It is expected behaviour?
No, it's a bug. During the first break a file was written that shouldn't have been written. Happens in specific circumstances.
Thank you for the testing! | |
|
|
|
... I've installed new Windows build, hopefully it'll fix the program.
Look like previous version was erroneously built with AVX instructions, which are not supported by your CPU.
Thank you stream, new Version 1.02 working so far without crashing. | |
|
streamVolunteer moderator
Project administrator
Volunteer developer
Volunteer tester
Send message
Joined: 1 Mar 14
Posts: 1022
ID: 301928
Credit: 543,195,386
RAC: 1
|
|
It's time to enable multithreading!
It was a surprise even for me, but FFT method used for factorial and primorials is not only slower that standard Proth / Riesel tests, but also consumes two times more cache memory!
The formula for factorials / primorials is:
Cache_usage = FFT_SIZE * 16
It means that standard desktop CPUs with 6MB or 8MB cache should switch to mulithreaded mode. For few days, two threads should be enough, than it'll better to use 4 threads when FFT size exceeds 256KB. I'll keep you informed when it happens.
Multithreading is available using app_config.xml. It's located in project directory - <boinc data directory>/projects/boincvm.proxyma.ru..../app_config.xml. It should be plain text file with following content:
<app_config>
<app_version>
<app_name>prst</app_name>
<cmdline>-t 2</cmdline>
<avg_ncpus>2</avg_ncpus>
</app_version>
</app_config>
If the file is already there with other setting, just add <app_version> block (without <app_config> tags). Change number of threads (2) to other value if necessary. Save the file and reload configuration files in Boinc client or restart the client.
| |
|
streamVolunteer moderator
Project administrator
Volunteer developer
Volunteer tester
Send message
Joined: 1 Mar 14
Posts: 1022
ID: 301928
Credit: 543,195,386
RAC: 1
|
|
New version of application (1.03) is installed for both Windows and Linux. It should fix an issue with restart from less efficient (older) checkpoint.
| |
|
HonzaVolunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 15 Aug 05
Posts: 1949
ID: 352
Credit: 6,014,223,837
RAC: 1,553,618
|
|
As candidates are growing (now 224K FFT), MT is going well into play and scales better.
Running 2x3 (2 tasks with -t 3) on Intel 9700K gives good performance.
Leading edge on PRPNet has 384K FFT so 6MB CPU cache is the edge.
Does not scale beyound 4 threads and it's not TDP or temp limit.
No overheating test there as I use to see with LLR2.
5950X works better in MT regard.
Largest known has 768K FFT hence CPU with 12MB CPU cache should be good but it's too much for i7 9700K to get is really busy.
Anyway, did a lot of Ctrl+C, changing number of threads and restarts, no complaning there.
____________
My stats | |
|
|
|
|
You can also try -spin <num_threads>, which enables spinwaits on all threads. Scales better, but consumes all CPU. Can be used only on dedicated crunchers. | |
|
HonzaVolunteer moderator
Volunteer tester
Project scientist
Send message
Joined: 15 Aug 05
Posts: 1949
ID: 352
Credit: 6,014,223,837
RAC: 1,553,618
|
|
Hmm, interesting.
For i7 9700K and 224 FFT, -t 4 - spin 2 looks like optimal combanation.
Gives best performace and not highest power comsunmption.
More threads is slightly slower and more power hungry.
Similar for FFT512K with different -t and -spin values.
Should -spin be smaller than number of threads as general rule?
Or is it architecture specific?
EDIT: AMD 5950X.
It rocks with -t 8 -spin 8
____________
My stats | |
|
|
|
|
Yes, spin can be less than <threads>, but I don't know why it can be beneficial. Usual values are either:
0 no spinwaiting;
1 only the main thread is spinwaiting for helper threads. When they're done, the main thread immediately continues with its operations while the helper threads go to sleep using OS calls;
<threads> all threads are spinwaiting. When the helper threads are done, they do not sleep but are waiting for the next job to do.
1 is default. | |
|
streamVolunteer moderator
Project administrator
Volunteer developer
Volunteer tester
Send message
Joined: 1 Mar 14
Posts: 1022
ID: 301928
Credit: 543,195,386
RAC: 1
|
|
FFT has grown again. Now optimal mulithreading setup is 3 or 4 cores, depending on your CPU.
| |
|
BurVolunteer tester
Send message
Joined: 25 Feb 20
Posts: 515
ID: 1241833
Credit: 414,225,485
RAC: 41,372
               
|
|
This is exciting news, especially that multithreading is now possible. Looking forward to it going live, the last discovery for primorial + 1 has been 21 years ago...
Btw, there was quite some outside-Primegrid work on Factorials in the last year. Will PG doublecheck those ranges?
____________
1281979 * 2^485014 + 1 is prime ... no further hits up to: n = 5,700,000 | |
|
tngSend message
Joined: 29 Aug 10
Posts: 465
ID: 66603
Credit: 45,703,545,082
RAC: 23,212,924
                                               
|
This is exciting news, especially that multithreading is now possible. Looking forward to it going live, the last discovery for primorial + 1 has been 21 years ago...
Btw, there was quite some outside-Primegrid work on Factorials in the last year. Will PG doublecheck those ranges?
I believe that that is being done as part of the PRST testing. On http://boincvm.proxyma.ru:30080/test4vm/ , under 'New application: "PRST testing"', it says"As a first step, we'll double-check PrimeGrid factorial and primorial PrpNet work, testing PRST itself and server-side scripts.
____________
 | |
|
|
|
|
Hi there!
I see a bunch of arithmetic errors almost in each WU. Is this ok or I have to check my hardware?
Arithmetic error, restarting at 2.6%.
Arithmetic error, restarting at 8.2%.
Arithmetic error, restarting at 23.3%.
Arithmetic error, restarting at 27.1%.
Arithmetic error, restarting at 28.1%.
Arithmetic error, restarting at 34.3%.
Arithmetic error, restarting at 45.7%.
Arithmetic error, restarting at 48.8%.
Arithmetic error, restarting at 64.6%.
Arithmetic error, restarting at 69.7%.
Arithmetic error, restarting at 70.2%.
Arithmetic error, restarting at 76.5%.
Arithmetic error, restarting at 81.2%.
Arithmetic error, restarting at 85.3%.
Arithmetic error, restarting at 90.1%.
Arithmetic error, restarting at 92.1%.
Arithmetic error, restarting at 94.2%.
Arithmetic error, restarting at 94.2%. | |
|
Nick
Send message
Joined: 11 Jul 11
Posts: 2216
ID: 105020
Credit: 8,133,261,170
RAC: 1,329,340
                   
|
Hi there!
I see a bunch of arithmetic errors almost in each WU. Is this ok or I have to check my hardware?
Arithmetic error, restarting at 2.6%.
Arithmetic error, restarting at 8.2%.
Arithmetic error, restarting at 23.3%.
Arithmetic error, restarting at 27.1%.
Arithmetic error, restarting at 28.1%.
Arithmetic error, restarting at 34.3%.
Arithmetic error, restarting at 45.7%.
Arithmetic error, restarting at 48.8%.
Arithmetic error, restarting at 64.6%.
Arithmetic error, restarting at 69.7%.
Arithmetic error, restarting at 70.2%.
Arithmetic error, restarting at 76.5%.
Arithmetic error, restarting at 81.2%.
Arithmetic error, restarting at 85.3%.
Arithmetic error, restarting at 90.1%.
Arithmetic error, restarting at 92.1%.
Arithmetic error, restarting at 94.2%.
Arithmetic error, restarting at 94.2%.
It ain't great but if you get a true result at the end - then it is satisfactory.
Check temperatures.
Don't overclock memory, or CPU, or GPU.
Make sure your PSU is up for it.
Memory is the most frustrating of things to diagnose. | |
|
|
|
|
We've approached FFT boundary, some roundoff errors are expected. PRST makes in-memory checkpoints every 20k iterations, so very little progress is lost after restart. Once all tasks switch to the next FFT, arithmetic errors will be gone (until next boundary). | |
|
|
|
|
These things are starting to get big.
I was running 4 core multi-thread per work unit and it was taking them 4 to 4 1/2 hours on a Ryzen 9 5900x and a Ryzen 9 7900x.
I have had to increase to 6 then 8 cores as my computer downloaded 119 of these things and times blew out to 9 hours so upped to 8 cores and times came back down to 4 hours 30 minutes to 4 hours 40 minutes.
I still have 87 to do with a deadline of 5 days from now, will see how I go.
Conan
____________
 | |
|
|
|
|
The numbers are larger than mega already. And not that many of them left, we're getting close to the leading edge. | |
|
|
|
|
Can you pin thread with Invitation code at https://www.primegrid.com/forum_forum.php?id=75 so others can find it more easily? | |
|
|
|
|
Less than 10k factorials of both types left!
https://boincvm.proxyma.ru:30443/test4vm/user_profile/llr2_status.html
The DC is almost over. | |
|
Post to thread
Message boards :
Number crunching :
Beta testing of PRST app is open on "GFN Server" |
