New Multi-Threaded Pi Program - Faster than SuperPi and PiFast

Quote:

---

Originally Posted by Sandon

Performance test is looking better with a new raid controlller.. hopefully should speed up my 13.3 trillion calculation by quite a bit I hope:

Code:

---

Sequential Write:        1.59 GB/s
Sequential Read:          1.77 GB/s
Threshold Strided Write:  864 MB/s
Threshold Strided Read:    881 MB/s

Overlapped VST-I/O Ratio: 0.779955

Notes:

  - The overall I/O speed is unable to keep up with the CPU(s).
    The I/O throughput is 1.28213x slower than the CPU throughput.
    Large computations will be significantly slowed down by disk access.
    I/O bandwidth can be increased in a number of ways:
      - Add more drives in parallel. This is the obvious way.
        Many machines have 4 or more drives just to run this program!
      - Defragment the drives.
      - Use empty drives. Empty and freshly formatted drives perform best.

  - Your threshold non-sequential I/O bandwidth is very high.
    This may cause sub-optimal algorithm selection for large computations.
    The optimal ratio between sequential/non-sequential I/O is about 3 to 1.
    It is recommended to decrease the "Min I/O Size" setting and re-run
    this benchmark.

---

I lol'd a bit during the sequential read test for a while it was > 2 GB/sec and I saw what I think is an easter egg =)

Code:

---

Sequential Read:          2.01 GB/s  WTF?!?!

---

The WTF?!?! was in blue.

---

Yeah, that's an easter egg. :) I should probably raise the threshold for that. I set it about 2 years ago and it's clearly becoming too easy to hit 2GB/s.
Back then, it was only possible via ram drive.

Let me know when the 13.3 trillion digit computation is getting near the end. (if you're still running it) Assuming no one beats you to it, it's going to be a new world record.
But to qualify as a new record, it needs to be verified. So at some point, somebody (either you, me, or someone else) will need to kick off the BBP program. That will probably take about a week on my spare machine (i7 920 @ 3.5 GHz).

Quote:

---

Originally Posted by NEOAethyr

I didn't even know you could disable avx...
Though I notice aida64 is saying it's disabled as well, but honestly I think it's just a screw up, It's probably working anyways.
There's alot of odd things here and there that don't work on my systems, past and present.
Cpu load, all sorts of perf counters and so on.
Anyways I'm gonna try this bcdedit mod to see if I can force avx to enable or whatever.

...

Anyways I'm off to play with bcdedit.
Though I doubt aida64 will beable to tell if it's working either way.


Update:
Ok well, forcing avx to enable doesn't work.
Avx is just isn't working on my setup lol, got the os, cpu and etc.
I apparently gutted avx along with float-16 and so on without realizing it, I wouldn't of known with my older amd cpu at the time.
I didn't think there was another say 80% perf boost just waiting for me lol.
I got a fresh os to the side I haven't finished setting up, no tweaks.
I planned on getting around to fixing up my tweaks for x64 win7 but just haven't gotten around to it other then installing windows and calling it quits for the time being, until now anyways.
Sigh I don't even wanna use windows lol, but I don't have the free space right now for linux...

I should try re-stressing my cpu with avx enabled a little later on, I thought I was using it but apparently not...
Linx went from 90 gflops to 160 gflops so apparently not lol.

---

I've updated the program to detect if the OS has AVX enabled. And if it isn't, it will spit out a warning and fall back to SSE3 or SSE4.1.
I haven't been able to test it since all my OS's are fully upgraded and I don't want to mess with bcdedit nor install another OS. But it *should* work fine.

This will go out in the next version (v0.6.5). But that might take a while.

Quote:

---

Originally Posted by poke349

Yeah, that's an easter egg. :) I should probably raise the threshold for that. I set it about 2 years ago and it's clearly becoming too easy to hit 2GB/s.
Back then, it was only possible via ram drive.

Let me know when the 13.3 trillion digit computation is getting near the end. (if you're still running it) Assuming no one beats you to it, it's going to be a new world record.
But to qualify as a new record, it needs to be verified. So at some point, somebody (either you, me, or someone else) will need to kick off the BBP program. That will probably take about a week on my spare machine (i7 920 @ 3.5 GHz).

---

I still am running it. Summing is currently been at 29% the last few days so definitely making progress. I guess I should let you know when summing is nearly 100%?

Quote:

---

Originally Posted by poke349

I've updated the program to detect if the OS has AVX enabled. And if it isn't, it will spit out a warning and fall back to SSE3 or SSE4.1.
I haven't been able to test it since all my OS's are fully upgraded and I don't want to mess with bcdedit nor install another OS. But it *should* work fine.

This will go out in the next version (v0.6.5). But that might take a while.

---

Okeydoke, I'll test it when you release the new ver, just give me a heads up in the release notes or something.
I'll install a new os and purposely screw it up just for this :up:.

Quote:

---

Originally Posted by Sandon

I still am running it. Summing is currently been at 29% the last few days so definitely making progress. I guess I should let you know when summing is nearly 100%?

---

Yeah. It'll take more than a week to run all the stuff after the series anyway. So you can wait until after the series is done.
Make sure you have hexadecimal digits enabled. Since that's needed for the verification.

Quote:

---

Originally Posted by NEOAethyr

Okeydoke, I'll test it when you release the new ver, just give me a heads up in the release notes or something.
I'll install a new os and purposely screw it up just for this :up:.

---

Don't hold your breath though. v0.6.5 isn't close to ready yet. I can't even give a rough date yet.
The fix for the AVX detection is actually a part of a complete revamp of the entire CPU dispatching logic.

Quote:

---

Originally Posted by poke349

Yeah. It'll take more than a week to run all the stuff after the series anyway. So you can wait until after the series is done.
Make sure you have hexadecimal digits enabled. Since that's needed for the verification.

---

I assume this means I have hexadecimal digits enabled as it says hexadecimal digit counts?

Code:

---

Constant :  Pi
Algorithm:  Chudnovsky Formula

Decimal Digits    :  13,300,000,000,000
Hexadecimal Digits:  11,045,410,915,501

Threads:    32
Mode  :    Swap Mode

Start Time: Thu Mar 13 03:51:59 2014

---

Or do I need to do something special? Can I enable it after I started the computation?

Quote:

---

Originally Posted by Sandon

I assume this means I have hexadecimal digits enabled as it says hexadecimal digit counts?

Code:

---

Constant :  Pi
Algorithm:  Chudnovsky Formula

Decimal Digits    :  13,300,000,000,000
Hexadecimal Digits:  11,045,410,915,501

Threads:    32
Mode  :    Swap Mode

Start Time: Thu Mar 13 03:51:59 2014

---

Or do I need to do something special? Can I enable it after I started the computation?

---

Yep, you got it. Once the computation is done, you'll need to use the Digit Viewer to read the last few 50 - 100 hexadecimal digits.
The program lets you disable the hexadecimal digits. I just wanted to make sure you had them enabled.

The BBP program will look something like this:

http://www.numberworld.org/misc_runs...ex_10_048t.jpg

The digits here will need to match the ones from the main computation. If they don't, then the computation is wrong.

If they don't match, I'll run the secondary BBP formula to help determine who's correct and who's wrong. Fortunately, this is has never happened before. All 3 of my runs with Shigeru Kondo (5, 10, and 12.1 trillion) had everything match at the end.

Just a small heads up. I've just released version v0.6.5.

This version adds optimizations for Haswell processors. It will use AVX2, FMA3, and BMI2 instructions.
I've also noticed that because of these optimizations, the program runs significantly hotter than before (on Haswell).

So please take care if you intend to run the stress-tester or any computation larger than 250 million digits on a Haswell processor.

Cool beans, I'll test it within a few days.

Offtopic.
I don't a working vbox on any of my os'es, seems if you install it twice on either linux or windows (on linux it screws dkms), it fails.
I gotta reinstall them both anyways but linux will take a bit to redo (I've got alot of house keeping todo on it to make sure I can script the next install even better, without killing vbox again...).

I think I figured out what disabled avx on my machine a while back.

Code:

---

;?, Screws over AVX and Float-16 conversion instructions when removed
;[-HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\services\hwpolicy]

---

I may just run that on my current os and see if it screws it over.
I wanted to run it on a vm though.
You just gotta give me a few, I just got ff v29.0.1 working halfway decently in win this morning (other then the fact I can't toggle js on the fly anymore wtf...).

Cool I can verify this new ver fixes the issue when os'es disable avx.
It gracefully falls back to a diff exe and runs fine.
Not that I need it anymore lol (I did after all end up having avx disabled by accident...), but still...

Code:

---

This processor is capable of running 256-bit AVX instructions.
But the OS is either too old or has disabled this capability.
To achieve maximum performance, you must have proper OS support.

Defaulting to the fastest compatible binary...

---

Quote:

---

Originally Posted by NEOAethyr

Cool I can verify this new ver fixes the issue when os'es disable avx.
It gracefully falls back to a diff exe and runs fine.
Not that I need it anymore lol (I did after all end up having avx disabled by accident...), but still...

Code:

---

This processor is capable of running 256-bit AVX instructions.
But the OS is either too old or has disabled this capability.
To achieve maximum performance, you must have proper OS support.

Defaulting to the fastest compatible binary...

---

---

Neat! Glad to hear that it worked. And thanks for going out of the way to test this for me. I really appreciate it!

Thanks for this! I've been using it to stress test my AMD Athlon x750k. Every other test I throw at it passes but the HNT test of Y-Cruncher. It represents CPU+Cache+Memory...but what to tweak to get it to pass this test? :p:

Quote:

---

Originally Posted by Dodgexander

Thanks for this! I've been using it to stress test my AMD Athlon x750k. Every other test I throw at it passes but the HNT test of Y-Cruncher. It represents CPU+Cache+Memory...but what to tweak to get it to pass this test? :p:

---

No idea. And there's no real way to tell what it breaks on. If none of the other tests reveal the problem, then I guess it's trial-and-error with the OC settings.

Yeah fft and hnt are the ones that always fails with mem and cpu it seems.
For now I'm sticking with all the tests, but I'll keep and eye on it.


I have a req...
Last night during testing I thought of something.
I'de like an option to halt on error in the stress test mode.

Reason, I want it to halt so I can figure out the prob, I'm not trying to hammer it in hopes of breaking it in.
Though that is an option..., I'm not running sub ambient at the moment.

So I would really like to see it as an option to halt it or leave it as the default, retry mode.
I know if it gets bad enough it will halt, but I'm looking for it to halt on any error.

That way I can sleep, and if it's unstable at all, the error will be right in front of me when I wake up.
For example last night, I went to sleep right after the 1st set of tests (I think there was 4 of 7 passes each).
I thought it possible that it could fail any min, a lower value for that mem timing caused an error within mins.
Fortunate for me it passed 8hrs :).

The idea anyways was if it did have an option, I could sleep soundly without having to watch it for an hour.
It wouldn't keep stressing it...

...
Putting it as an option, making it retry errors may be handy if you have your ram at really low temps.
I would like them both...

I would like to say though your program is the best for cpu testing.
Enabling avx improved the low vcore detection quite a bit.

Oh and eh..., is it possible to use the /binaries/*.exe directly with the cmd line (cmd tail) ?
I'm thinking that's what the front end does anyways...

So I posted an answer here earlier, but it got lost through the server downtime...

Quote:

---

Originally Posted by NEOAethyr

Yeah fft and hnt are the ones that always fails with mem and cpu it seems.
For now I'm sticking with all the tests, but I'll keep and eye on it.


I have a req...
Last night during testing I thought of something.
I'de like an option to halt on error in the stress test mode.

Reason, I want it to halt so I can figure out the prob, I'm not trying to hammer it in hopes of breaking it in.
Though that is an option..., I'm not running sub ambient at the moment.

So I would really like to see it as an option to halt it or leave it as the default, retry mode.
I know if it gets bad enough it will halt, but I'm looking for it to halt on any error.

That way I can sleep, and if it's unstable at all, the error will be right in front of me when I wake up.
For example last night, I went to sleep right after the 1st set of tests (I think there was 4 of 7 passes each).
I thought it possible that it could fail any min, a lower value for that mem timing caused an error within mins.
Fortunate for me it passed 8hrs :).

The idea anyways was if it did have an option, I could sleep soundly without having to watch it for an hour.
It wouldn't keep stressing it...

...
Putting it as an option, making it retry errors may be handy if you have your ram at really low temps.
I would like them both...

I would like to say though your program is the best for cpu testing.
Enabling avx improved the low vcore detection quite a bit.

---

Yeah, that's an easy enough feature request. Expect to see it in the next version. (Thanks for the suggestion. ;))

http://www.numberworld.org/y-crunche...6.6_stress.png

Quote:

---

Oh and eh..., is it possible to use the /binaries/*.exe directly with the cmd line (cmd tail) ?
I'm thinking that's what the front end does anyways...

---

Yeah, you can run them directly. And you're right that the front-end simply calls the other ones. Version 0.6.6 will also have real command line options that can be passed into either the front-end (y-cruncher.exe) or the individual binaries.