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

[poke349]

It was trashing like crazy for 2.5b. I wasn't running anything else in the background other than my normal apps. I expected 2.5b to be hard on my system though.

Interesting. At least on my machine, even when I have more than 500MB of stuff open, the program will page it out enough to let the program have the full 11.5 GB for the entire run.

Win7 change in memory manager?

Anyone want to try and confirm this?

yes it is.

your software is very good.

If i run Multi-Threaded Pi with 350.000.000, i run well Prime95 small FFT in 30 minutes without any error.

My PC Q9550 @ 4GHz.

P.S : Sorry my english.

thx

Did you verify the 350m? Only the benchmarks and the two of the batch mode options will verify to see if the digits are correct. The "Custom Compute" option does NOT since there are no checksums for them.

Just because it finishes without warnings doesn't mean it finished without any errors.


The stress test, on the otherhand, DOES verify to see if everything is correct.

The way it works is that it cycles through all the major constants that the program can compute. For each constant, it runs two computations using different algorithms. Then it matches them to see if they are correct.

It uses two threads that run completely independently of each other. Because of "dips" in the CPU usage for each computation, one isn't quite enough to keep everything busy. But two does pretty well - at least up through 8 cores.

[KB24LA]

lame result but still...

[poke349]

lame result but still...

nah... That looks about right for your processor/frequency.



Something interesting here:
http://www.itocp.com/bbs/thread-35879-1-1.html

The guy did a string table hack on the binaries. It'll print correctly only if you have your regional settings set to read ascii in some specially encoded way that works for Chinese characters.
And it isn't complete. Some of the other "secondary features" aren't translated at all.


Which gives me an idea. For next the release (v0.4.3), how about I link to an external .ini settings file that has the full string table in unicode?
That way it can be easily translated to any language by simply editing the .ini string table.


It's kinda drastic. But, I plan on implementing a very aggressive anti-tamper protection into v0.4.3 that will pretty much block any changes to the binary.
Which would make translation (via string-table hack) impossible unless the protection itself is broken. So the only way to allow any sort of language support is to move the strings out of the binary.


Any other ideas?

[Ket]

Heres my score with Q8400 @ 3.6GHz, 485FSB. RAM @ 1165MHz

[Peen]

Little bit faster 4.6ghz, 32m


and 25k


and 50k

[Chumbucket843]

dude your records are on wikipedia.

[Peen]

Mine are? Link me I know I can go faster

[Chumbucket843]

Mine are? Link me I know I can go faster

you probly set a single socket speed record but poke set the record for most digits calculated on several constants. i couldn't find some of the other ones but they need to updated.
http://en.wikipedia.org/wiki/Euler-Mascheroni_constant
http://en.wikipedia.org/wiki/Catalan's_constant
looks like kondo beat you on e!
http://en.wikipedia.org/wiki/E_(mathematical_constant)

[Peen]

Where do you see the times at? BTW cool links, wish I was smart enough to understand math like that!

[poke349]

dude your records are on wikipedia.

If anyone is still wondering what all that ram was for...

you probly set a single socket speed record but poke set the record for most digits calculated on several constants. i couldn't find some of the other ones but they need to updated.
http://en.wikipedia.org/wiki/Euler-Mascheroni_constant
http://en.wikipedia.org/wiki/Catalan's_constant
looks like kondo beat you on e!
http://en.wikipedia.org/wiki/E_(mathematical_constant)

Here's the one you missed:
http://en.wikipedia.org/wiki/Ap%C3%A9ry%27s_constant

There aren't any pages for the two Natural Logs.


As for e... well...

Here's my excuse:

Using the Basic Swap Mode, all computations of N digits require a little more than 2N memory.
So with 64GB of ram, the largest computation I can do is 31 billion digits.
So until I write an Advanced Swap Mode (the thing that PiFast and QuickPi have), I'm stuck at 31 billion digits with my "puny" 64 GB of ram.

32GB would've been enough to break the records. But at the time we built the rig (September 2008), the memory requirement was expected to be 4N (I hadn't implemented it yet).
So 64GB was needed to break 10 billion. (which was what the records were at the time.)

Then in December, I realized a 2N algorithm - which effectively allowed us to hit 31 billion.

Any higher, and multiplication will no longer fit in ram. So it will require a completely different scheme for operating on HDs.

Where do you see the times at? BTW cool links, wish I was smart enough to understand math like that!

I've updated the "Fastest Times" section on my site with your new numbers... Insane 4.6 GHz...

[rge]

Run with my i950

[poke349]

Woah

New single-socket records again!!!
EDIT: Is it safe to assume that's on water?

Single-socket is closing in on Dave's Gainestowns for 25m...


Just curious, how stable is that OC? I noticed that turbo is disabled.



And if anyone has screenies of a failed benchmark. I'm curious to see them.
Whether it fails because the digits don't match at the end, or if it catches and corrects an error....

Be kind of interesting to see what the most common non-crash/BSOD failure is.

[rge]

EDIT: Is it safe to assume that's on water?

Just curious, how stable is that OC? I noticed that turbo is disabled.

Yep water, pic is in sig. Turbo on GB board is just 24 multi all the time, 23 multi OC's better, so turbo is off.

4.6ghz is stable enough to run prime blend 20 mins without crashing, some with bigger cojones primed 4.6ghz for 8hrs posted in the i7 database intel thread.

4.65 stable to run chess computer stress, posted there.

4.69, ran 25 and 50 several times, never crashed. 100 tried to run once crashed (bsod) about 3/4 way through. Did not try upping vcore more, probably wont until winter and can set rad in cooler temps. It is between 4.6 and 4.7 where stability markedly changes.

[Peen]

hmmm looks like I'll have to try 4.75ghz now hehe. Kinda board limited though but haven't really pushed too hard

[xman01]

did batch run on rig 1 and rig 2 on winxpx64
hope this is valid enough


rig2


rig1

[poke349]

did batch run on rig 1 and rig 2 on winxpx64
hope this is valid enough

Good enough... lol

I haven't been validating any of these anyways... As long as they "seemed right", good enough, lol.

I'll probably put in a hash for the batch benchmarks later on.


Been really busy lately... Not gonna have anytime until after GREs...

[rge]

downstairs with ~18-19C ambients got 100 to run, and little higher, 4.76ghz, 1.55vcore.

[dj883u2]

Run with my Core i7920@4210Mhz

[poke349]

Run with my Core i7920@4210Mhz

Woah... Is that voltage for real? Is that one of those 3845/3849 batches?

With 1.15v I can only get up to 3.9 GHz... and only marginally stable.
I need 1.225 to bench @ 4.2, 1.25 to "barely"pass 2G Pi, and 1.275 to get prime/linpack stable.


And just wondering, did you play with memory timings? I'm curious to see how sensitive the program is to timings.

[dj883u2]

Woah... Is that voltage for real? Is that one of those 3845/3849 batches?

With 1.15v I can only get up to 3.9 GHz... and only marginally stable.
I need 1.225 to bench @ 4.2, 1.25 to "barely"pass 2G Pi, and 1.275 to get prime/linpack stable.


And just wondering, did you play with memory timings? I'm curious to see how sensitive the program is to timings.

Yes, the voltage is real. The Batch is 3845B027.

What tests do you need?

[poke349]

Yes, the voltage is real. The Batch is 3845B027.

What tests do you need?

Ah... Something like different timings for 25m and something larger like 250m. And maybe with and without HT. But I shouldn't be asking for anything.

Threading overhead and randomness makes the small benchmarks somewhat unreliable as an indicator of program efficiency.
HT has an effect of "hiding" latencies because when one thread stalls for memory access, the other thread gets the whole core to itself...


I haven't gotten around to trying it myself because my rig is usually busy with stuff that I can't pause.
Obviously this program is very memory intensive. And I know for sure that Core 2 bandwidth is bottlenecking. But I've never been able to gauge the effect of latencies.

Timing sensitivity, I would guess, is a good indicator to whether the program will scale well on EX servers.

[Chumbucket843]

you could buy vtune for like 1000 dollars.
http://software.intel.com/en-us/intel-vtune/

[poke349]

you could buy vtune for like 1000 dollars.
http://software.intel.com/en-us/intel-vtune/

I don't take optimization seriously enough to justify that kind of expenditure.

[Peen]

Was about to do 4.7ghz+ but then my board died again! damn u dfi

[maxgull]

thought i'd join in..