JJSieve Discussion

[pixl97]

What is jjsieve exactly, I've been experimenting with it the last few days and it seems to work on my athlon64

[Matt]

I searched jjsieve and found a thread discussing some testing that happened a month or so ago, any update on when this client will be availiable? If it offers a significant speed boost (as seems to be suggested) I'm sure we'd all be keen to see it!

[Greenbank]

Have a look at the bottom of the 1st post in this thread. It's attached there.

[shauge]

Does it being posted here mean that we can use it for regular sieving now? I guess it is not "released" yet, but are we confident it will find all the factors "proth_sieve" will find?

[jasong]

Jjsieve is very dependable and, I'm told, will even find factors that proth sieve might miss.

[Matt]

How about clients for other operating systems? I run 3 of my sievers on FreeBSD, also is there any documentation related to this?

[hhh]

Related to this jjsieve discussion, I got some questions.
In the above zip file, one finds jjsieve, jjsievecmov5, jjsievecmov6, jjsievecmov7. What is the meaning of this?
What about sse2 (there was such a jjsieve a while ago)?
And is there a Linux binary?
Can the joint sieve subproject take advantage of the impovements, too?

That's about it, for the moment, I guess... Yours, H.

[KWSN_Dagger]

There is an SSE2 version out there, but the one listed in this forum is old (EDIT not any more!). Joe O can explain more than I.
jjsieve - no CMOV, no SSE, no SSE2
jjsieveCMOV5 no SSE, no SSE2
jjsieveCMOV6 SSE, no SSE2, small to medium cache
jjsieveCMOV7 SSE, no SSE2, medium to large cache
jjsieveSSE2

[shauge]

I tried the CMOV7 version on my AMD64, I got a 5% speed increase. Nice to see the ETA shorten by 10 hours :) but I look forward to the SSE2 version.

[shauge]

The SS2 version is now added. I got about 12% speed increase using that on my AMD64 and I got about 15% speed increase on my AMD MP2400 using the CMOV5 version. I think this is great!

[hhh]

Linuxlinuxlinux. Possible? I can compile it if necessary. P4, 3.06GHz. H.
WineWineWine! Grin! <--------------------Edit

[cedricvonck]

Hi

I have been crunching some range and found about 10+ factors.
Are the factor well received in the DB?
And what about the scoring stuff?

Can we use jjsieve as production stable?

Any update will be appreciated.

Thank you

[Greenbank]

Scores (updated twice a day):-

2006: http://www.aooq73.dsl.pipex.com/2006/scores.htm
All time: http://www.aooq73.dsl.pipex.com/scores.htm

Probably best to look at just the 2006 scores...

Note that only factors for n < 20M are scored at the moment.

Meanwhile, the disk in one of my Athlons has died. Was about 97% through a 1T range on it so it shouldn't take long to finish off but will be delayed a week or so. It now becomes a diskless host and I get another bunch of magnets to play with.

[cedricvonck]

Greenbank

Thank you for the heads up.
But how do you explain this
http://mattford.org.uk/sob/userinfo.php?user=84

1022500 1022510 0 Complete March 21, 2006 March 15, 2006
1023060 1023110 0 Complete March 21, 2006 March 19, 2006
1029700 1029750 2 Complete April 11, 2006 March 31, 2006
1033150 1033200 0 Complete April 11, 2006 April 8, 2006
1034120 1034170 4 Complete April 22, 2006 April 14, 2006
1036000 1036050 2 Complete April 22, 2006 April 19, 2006
1044800 1044850 4 Complete April 25, 2006 April 24, 2006
1046800 1046850 2 Reserved April 22, 2006 April 27, 2006
1049700 1049800 2 Reserved April 24, 2006 April 28, 2006
1101000 1101005 ? Complete March 21, 2006 Unknown
1102000 1102100 ? Complete March 21, 2006 Unknown

Total factors found = 16

The things don't add up....

[Matt]

The factors found on the SoB reservation site is simply a note that the user who runs the test can make, it's completely unrelated to the sieve scoring stats and is probably best to be ignored.

Anyone mind if I make a seperate thread to discuss jjsieve as it's kind of off topic here...?

[hhh]

Anyone mind if I make a seperate thread to discuss jjsieve as it's kind of off topic here...?
Go ahead. No askingneeded, I think. H.

[Matt]

Right, so is this officially released? Stable? Cross-platform?

[pixl97]

Just tried the JJsieveSSE2, seems a good bit faster then JJsieveCMOV7. My numbers went from about 540kp/s to over 580kp/s on my athlon 64 3000+

Now all i need is a linux version of it!

[KWSN_Dagger]

Just tried the JJsieveSSE2, seems a good bit faster then JJsieveCMOV7. My numbers went from about 540kp/s to over 580kp/s on my athlon 64 3000+

Now all i need is a linux version of it!
I'm averaging 680 kp/s with highs of 695.

[Matt]

I'm itching to get my hands on this, I've got an AMD64 3000+ on FreeBSD currently running at 481kps. Also an AMD Barton XP3000+, again on FreeBSD, running at 543kps. Basically I'd really like a FreeBSD client, I can offer SSH access if someone wants to compile one?

[KWSN_Dagger]

Also works fine for combined sieving with PSP. Speeds of 383 kp/s so far. Not bad when your going through 20k's. Big :clap: to Joe O for making this lil one run.

[Matt]

I've been running JJsieve with wine in FreeBSD and here's the basic summary of what I get, on my Athlon XP3000+, with CMOV7 I get about 530kp/s compared to a previous 540kp/s (so actually a reduction). However, on my AMD 64 3000+ with the SSE2 client I get a speed of about 570kp/s compared to a previous speed of 450kp/s, so that's about a 26% increase in speed! Good work!

[cedricvonck]

On my P4 2400 I am running @ 300kp/sec.
Only a "minor" increase with 25kp/sec but glad to have it.

BTW my Athlon 1 ghz (proth sieve) is running @ 275 kp/sec :slap:

[Joe O]

I've been running JJsieve with wine in FreeBSD and here's the basic summary of what I get, on my Athlon XP3000+, with CMOV7 I get about 530kp/s compared to a previous 540kp/s (so actually a reduction). However, on my AMD 64 3000+ with the SSE2 client I get a speed of about 570kp/s compared to a previous speed of 450kp/s, so that's about a 26% increase in speed! Good work!
Please try CMOV6 on both machines. You may be pleasantly suprised. CMOV6 outruns CMOV7 on many Athlons, even keeping up with SSE2 on some AMD64's. YMMV

[Joe O]

On my P4 2400 I am running @ 300kp/sec.
Only a "minor" increase with 25kp/sec but glad to have it.

BTW my Athlon 1 ghz (proth sieve) is running @ 275 kp/sec :slap:

You didn't state what version you were using. Please try both CMOV5 and CMOV6 on these machines and report the timings.

[cedricvonck]

Joe_O :

P4 2400 - JJSieve SSE2 => 300 - 301kp/sec
P4 2400 - JJSieve Cmov => unknown because range in progress
P4 2400 - Proth Sieve 0.42 SSE2 => 275 kp/sec

Athlon 1 Ghz - JJSieve Cmov => unknown because range in progress
Athlon 1 Ghz - Proth Sieve 0.42 CMOV => 275 kp/sec

[Matt]

Thanks for the advice, I've tried the other versions and they definitely run better.
AMD Athlon XP3000+ Barton:

Proth Sieve: 540 kp/s
JJSieveCMOV5: 580 kp/s
JJSieveCMOV6: 580 kp/s
JJSieveCMOV7: 550 kp/s

[Chuck]

At the risk of causing anyone to do an emergency run to the restroom.... :D


I am running a test across an entire range of P (from P=100G and up)

You will be pleased to know the client is doing FABULOUS.

Riesel and SoB are performing at 187 kp/sec and 940 kp/sec respectively.

I even am (shudder) testing using an Intel (Prescott) core... LOL
It is doing great.

Using 7 machines, I have a combined (summing the 'kp/sec') of 5940 kp/sec,
with the Intel in the suite of 7.

This is using ver 0.102a of the client.

Even a P4 Prescott getting 130 kp/sec on Riesel says something.. doesn't it?

When reporting back info to Joe or myself, please make sure (as joe pointed out) to let us know what version and any other pertinent info.

Chuck

[KWSN_Dagger]

With JJSieve SSE2 i'm getting 160kp/s on Riesel with my X2 4200 for a total of 320 kp/s.

On SoB I was running max 697kp/s for each core.

With PSP not combined ~ 450kp/s
Combined ~ 388kp/s

ver 0.102a

[cedricvonck]

Chuck,

One nice feature I liked to see ported from proth sieve to JJSieve (or whatever) is the following:

If you start the client up: JJSieve 157670 157700
I expect the client should interpret this as following:

pmax=1576700000000000
pmin=1577000000000000

Instead of JJSieve 1576700000000000 1577000000000000

Risks of doing starting the client up with the extra zeros, you can crunch a wrong range.

Is this possible?

BTW, who did the development of this new sieve client?
It is quite unclear for me.
I assume you, Greenbank, Vjs and Joe_O ?
Only for information purposes of course :D


RieselSieve : P4 - 3000

JJSieve SSE2 = 79kp/sec
Proth Sieve SSE2 = 60 - 64 kp/sec

Strange things, I expected a higher crunch rate... anyways thx

[Chuck]

The expansion of the -c command line argument to accept an 'exact' pmin and pmax was Joe's decision. I believe Joe can put it to rest, but given the desire to have a restart (after shutdown or your pc powerfails, etc)... you are restarting at an exact P... so why not be able to specify it from the start.

Also, the desire to do network task distributed sieving required the server to distribute exact Pmin, Pmax and receive Pcompleted as non-rounded numbers.

I understand your request for a more friendly interactive entry mechanism. How about putting in PRS -d -w -c 145260G 145360G (where the "G") puts up the extra zeros for you? This would mandate K=Kilo, M=Meg, G=Gig, T=Tera, P=Penta.. expressed as powers of 10^{3,6,9,12,15,etc} Is that ok?

As for who did the work??? Joe and myself. Vjs has been here in the background and is famous for being the silent type... I would not be surprised to find some of his work in here.

Plain and simple.... Joe gets 99% of the credit as far as I am concerned as he is the team leader and made sure everything got done. My job was to double check, experiment, verify things in 64 bit, etc. I did all the odd stuff. Joe is the Keystone to PRS's success in Windows as I see it. (He may say otherwise, but "" Joe, you know better.... LOL ""), I am the Linux/Unix side guy and the 64 bit guy. I know we watched each other's back in development, but I have to give Joe the credit for being the stronger of the two of us.... hands down.

There is no Mac code (Alex / Mark) in our code at all.

There are no plans at present to make the source open because we are simply 'static' for the moment. We both need a break. I have the complex task of mastering the new 64 bit stack frames going back to GCC... YES, I said GCC. Intel is our 32 bit compiler, but I MUST have GCC for Linux and BSD as well as 64 bit executables.

The code is already 64 bit native. There is ONE asm module in the whole thing and it is worth the effort in speed to keep it for now. It's very tiny, but critical to doing some lightning fast math.

With the changes we still have before us, this is the reason we won't be releasing the source yet... It will be complete and uniform first... Anything else could lead breaking the accuracy / stability / intelligence / reliability which has been put into the code. I am not saying anyone is a dummy, but it is VERY easy to break the code or (worse) slow it down or cause bad factors. We want this all rock solid before it all comes out.


Now, if you will forgive me... It's 5 am, time for breakfast, etc.... Lots to do today.

I will be aroudn later.


[[[[[[ Joe: Please do correct me if my slumbery mind has rambled and/or misrepresented anything ]]]]]]


Thanks
Chuck


/*edit -
Working on a 'process note' for all..... PRS SSE2 version, using 2.8 Ghz Prescott core.. 130+ on Riesel.. as compared to 79+ without using the 'tweaks'. I will report the stats for SoB as well when completed. *end edit */

[cedricvonck]

Thank you ! :rock::thumbs: :thumbs:

[Joe O]

The true and proper name for this program was to be CJJSieve, but a typo resulted in jjsieve. (Chuck, that was a typo wasn't it? and please explain to me again why we couldn't correct it.)
The next name we chose was PSR52 (PrimeSierpinskiRiesel) or was it PRS52(Prime Riesel Sierpinski or ProthSieve)?
At one time, I suggested (threatened?) SPQR but a guy name Caesar said that had already been used.
We've had a lot of people testing this program. Pepe, Lars, Bruce to name just a few. To all of you "Thank You". Our emphasis has been on reliability, and I think that we have achieved it. There is more to do, but for now, we are putting a hold on further development. We all need a break, and the immediate needs have been met.

[vjs]

Nope not me, I realized very early on that Joe and Chuck are far far beyond me when it comes to coding. If I can take any credit whatsoever it would be putting these two masterminds together.

[Chuck]

/* edit --- Comment withdrawn */

[hhh]

If I understand correctly, the question is about the name of the thing. THE NAME??? The least important thing at all, isn't it? But if you insist on my opinion, I prefer jjsieve, because it is easy to pronounce and remember, unlike PRS52 etc., and it looks better, too. For me jj stands for a revolutionary breakthrough, and is originary.
If he wants, Chuck can change his Alias to Juck, if he wants to.:jester:

Yours H.

[Joe O]

If he wants, Chuck can change his Alias to Juck, if he wants to.:jester:

Yours H.

Only if you change yours to JJJ. <G>

[Greenbank]

Just to clarify. I've had nothing to do with this sieve, Rogue and I have been working on porting proth_sieve to G5 PPC assembly for PowerMacs but this is completely separate to the work that Chuck and Joe_O have been doing for x86.

Good work gentlemen.

[ShoeLace]

The true and proper name for this program was to be CJJSieve, but a typo resulted in jjsieve. (Chuck, that was a typo wasn't it? and please explain to me again why we couldn't correct it.)
The next name we chose was PSR52 (PrimeSierpinskiRiesel) or was it PRS52(Prime Riesel Sierpinski or ProthSieve)?
At one time, I suggested (threatened?) SPQR but a guy name Caesar said that had already been used.
We've had a lot of people testing this program. Pepe, Lars, Bruce to name just a few. To all of you "Thank You". Our emphasis has been on reliability, and I think that we have achieved it. There is more to do, but for now, we are putting a hold on further development. We all need a break, and the immediate needs have been met.

:umm:
instead of SPQR.. perhaps you could call it PQRS as in lmo-PQRS-tuv
:lmao:

[royanee]

Linuxlinuxlinux. Possible? I can compile it if necessary. P4, 3.06GHz. H.
WineWineWine! Grin! <--------------------Edit

Wine is bad!!! At least in my experience:

430 kp/s top for JJSieveSSE2 under Wine
500 kp/s top for proth_sieve_sse2

Setup:
Linux 2.6, Athlon 64 3000+, 1 GB Dual Channel DDR
Currently sieving @ 1062.3 T

edit: transposed the 4 and the 3

[Joe O]

:umm:
instead of SPQR.. perhaps you could call it PQRS as in lmo-PQRS-tuv
:lmao:
SPQR was emblazoned on the standards of the Roman legions and was used by the Roman republic and the Roman empire. Caesar refers to Julius Caesar. It is the acronym for Senatus PopulusQue Romanus which roughly translates into The Senate and People of Rome, or the Senate and the Roman people This started to be used since a very early stage of the Roman republic, and later continued to be used during the Roman empire. As such, it appears in most of the famous monuments and documents. A fine example of this is the Arch of Titus built around 81 AD to honor Titus and his father the Emperor Vespasian. It is also used in Trajan's Column which was built in 113 AD to pay homage to Emperor Trajan.

For extra credit please identify: "Quis custodet ipsos custodes" and/or "Moriture te salutamus".

ps did you mean lmno-PQRS-tuv?

[ShoeLace]

S

ps did you mean lmno-PQRS-tuv?


yes.. i forgot the n :(

[Greenbank]

For extra credit please identify: "Quis custodet ipsos custodes" and/or "Moriture te salutamus".


My Latin is very rusty, but the first one is easy. "Who watches the watchmen?" (Or polices the police, or guards the guards). It's usually custodiet if I remember correctly.

Second one, hmm, Death, salute...ah yes, the gladiatorial greeting before a fight.. "We who are about to die salute you."

[Joe O]

My Latin is very rusty, but the first one is easy. "Who watches the watchmen?" (Or polices the police, or guards the guards). It's usually custodiet if I remember correctly.

Second one, hmm, Death, salute...ah yes, the gladiatorial greeting before a fight.. "We who are about to die salute you."
Correct on both, and the form of the verb as well.

[Joe O]

Wine is bad!!! At least in my experience:

340 kp/s top for JJSieveSSE2 under Wine
500 kp/s top for proth_sieve_sse2

Setup:
Linux 2.6, Athlon 64 3000+, 1 GB Dual Channel DDR
Currently sieving @ 1062.3 T

Unfortunately, comparing the "top" reported speed for proth_sieve will lead to inaccuracies. There were two loops in it, one for the processing and one for the reporting. Because of the relative size of the controlling variables, there was a 5 5 6 beat pattern in the reported speeds. you have to take the average of 16 consecutive reported speeds to get the true speed.

[ShoeLace]

there was a 5 5 6 beat pattern in the reported speeds. you have to take the average of 16 consecutive reported speeds to get the true speed.

thats a useful thing to know.. cheers

[royanee]

Unfortunately, comparing the "top" reported speed for proth_sieve will lead to inaccuracies. There were two loops in it, one for the processing and one for the reporting. Because of the relative size of the controlling variables, there was a 5 5 6 beat pattern in the reported speeds. you have to take the average of 16 consecutive reported speeds to get the true speed.


hehe, then I'll be exact:

n=49
348.47 with a standard deviation of 41.56
465.66 with a standard deviation of 18.99

If anything, the second set of number was under the higher load, because I let the JJsieve go off on it's lonesome, until I realized how slow it was going. When I started proth_sieve, I went back to using the computer, so the averages dropped from ~494 to the ~466 shown.

[Zahmekoses]

Hi there :)

Just one question from my side (or maybe reporting of a bug)...

I'm using JJSieveSSE2 under WinXP Home on an Mobile AMD Athlon XP64 4000+, sieving for Riesel Sieve, and noted that my jjsieve client does not seem to accept the "-i" switch.

When I start the program with "JJsieveSSE2.exe -i 1000000 -vv" it writes its progress every 67'000'000 into the RieselStatus.dat and not, as it is suggested by the switch, every 1'000'000. I've tried a bit around and altered the number behind the -i to numbers from 500'000 to 10'000'000, I even stopped using the -i switch, but everytime the client sieved approx. 67'000'000 before he gives me a line like "pmin=1634XXXXXXXXXXX @ 197kp/s".

Is it my fault that it does not work (maybe bad karma? ;o)) or have others also experienced this behaviour of the client?
It's not the end of the world, but I just wanted to point that one out :)

Regards,
-Zahme

[vjs]

I think Joe may have actually removed or reduced the reporting. There was a point where chuck was talking about the code and when to get the client to report a speed and p.

I'd test it with an sob dat first if it doesn't work I think that function was removed. Or at least it doesn't work as suggested.

I can't recall correctly but I think it had something to do with the prime generation??? Perhaps Chuck can step in and answer.

[Joe O]

Hi there :)

Just one question from my side (or maybe reporting of a bug)...

I'm using JJSieveSSE2 under WinXP Home on an Mobile AMD Athlon XP64 4000+, sieving for Riesel Sieve, and noted that my jjsieve client does not seem to accept the "-i" switch.

When I start the program with "JJsieveSSE2.exe -i 1000000 -vv" it writes its progress every 67'000'000 into the RieselStatus.dat and not, as it is suggested by the switch, every 1'000'000. I've tried a bit around and altered the number behind the -i to numbers from 500'000 to 10'000'000, I even stopped using the -i switch, but everytime the client sieved approx. 67'000'000 before he gives me a line like "pmin=1634XXXXXXXXXXX @ 197kp/s".

Is it my fault that it does not work (maybe bad karma? ;o)) or have others also experienced this behaviour of the client?
It's not the end of the world, but I just wanted to point that one out :)

Regards,
-Zahme
No, not your fault. The reporting interval has been changed to be in sync with the processing interval. If you increase the number to 100'000'000 you will see that it reports every 135'000'000 or thereabouts. The open question is how to change the description of the -i switch. should we change it to a multiple ie 1x, 2x, 3x, etc. or just drop it. Of all the people from all the projects that we talked to, only one was using it and he was happy with the new number. He had been increasing it to lower the disk activity. This interval also controls how often the pmin checkpoint is written to the status file.

[hhh]

I just had a (possibly useless) idea.
I have two instances of prothsieve running on my hyperthreaded P4. So, I had to split my range in two halfs.
I'm sure it is not difficult to implement BACKWARDS sieving; so one client could easily take advantage of a HT/Dual Core CPU. It starts just at the beginning and the end of the range at the same time.
Don't know if such a feature would be apreciated, though. With a switch, perhaps? Yours, H.

[Mystwalker]

I posed such an idea some months (or even years?) ago. IIRC, there was no result. The problem is that the code has to be rewritten to allow this, which will be a lot of hassel. :(

On a HT machine, I wouldn't use 2 proth_sieve instances anyway. Instead, one proth_sieve and some other client (e.g. PRPing or P-1 factoring) is more efficient in my experience.

[hhh]

On a HT machine, I wouldn't use 2 proth_sieve instances anyway. Instead, one proth_sieve and some other client (e.g. PRPing or P-1 factoring) is more efficient in my experience.

I believe this is an urban legend (which I helped to spread, I confess).
I did some tests some time ago, and 2xproth just doubles my output.
2xPRP gives half the speed for each, and proth+PRP doesn't change the speed of proth, but gives half the speed for PRP.
So, if I do one month proth+PRP, it is the same thing as half a month PRP (at double speed) and half a month 2xproth.
I tried to put two instances of P-1 together, too, and it didn't wark any better (this idea was btw my first post at mersenneforum).

Yours H.

[vjs]

I think the major advantage of running the sieve or PRP along side P-1 is the somewhat fixed low memory consumption. Generally allowing more memory for P-1 stage2 on the second processor.

[Mystwalker]

I did some tests some time ago, and 2xproth just doubles my output.

That's odd - my results were somewhat different. I don't remember exact figures, but using two instances increased throughput by 40-60% in my testings with my P4.
When running a single proth_sieve and then starting e.g. PRPing, sieving speed didn't halve, but dropped by maybe 30-40%...

Unfortunately, I can retest it now, because the P4 system isn't stable anymore since 2-3 weeks. When stressing the CPU, it reboots after maybe 30 minutes. :(

[hhh]

When I'm running one proth-sieve and starting something else, the speed of proth isn't affected. So I have the choice between another proth at the same speed or one prp at half speed.
The latter doesn't change from sieving 2 times and then only PRPing.
With P-1, I didn't try out.
Yours H.

[Mystwalker]

When I'm running one proth-sieve and starting something else, the speed of proth isn't affected.

That's interesting.
I'm very confident you know the difference, but I have to ask in order to be sure:
You really have a HT system and no dual-core, don't you?

[hhh]

I know the difference, but: it's my office pc, so I didn't check quite well. I remember enabeling HT in the bios, but that doesn't mean anything.
Linux tells me:
proc 0
Genuine Intel
family 15
model 4
P4 3.00GHz
stepping 1
cache size 1024KB
etc.
in the flags there is "ht"

Some proc freak might warp in here.
H.

[dudlio]

I just got a dual-core Athlon so I'm interested in running 2 instances and getting my cpu usage up to 100%. I reserved a second range, copied my sieveA folder to a new location, cleared out all the old config data and input the new range.

Then I fire up JJSieveSSE2.exe. My cpu usage jumps to 100% with 2 instances running BUT the second instance is still using the old range!

Any tips on running 2 ranges simultaneously? It's not working for me.

[dudlio]

Aha, the SieveB folder is for dual-core procs. Nice, it works. I can sieve@1600kps now. That's sick...

[Death]

Hi everybody )))
I'm back now!!!!

Greetz 4 all who remember me, howdo folks?

And now the problem ((((

cant start jjsieve under 2k pro
both sse and cmov version show this

Popup window: JJsieveSSE2.exe - application error : command at address "0x77f87eeb" call the memory at address "0x00000034". memory can't be "read".

[Death]

OMG

i forget to put sob.dat at directory and it just CRASH!!
now it works... (SSE version crash too, instead of saying smth like "no sse - i'm quit")

[Death]

and other question

does it goes right when shows 180k at p4 3.2 GHz ?
combined sieve 666000g range

[hhh]

Yes, that looks right to me. However, if you have hyperthreading enabled, you should run a second sieve client. Just make another folder, and run it on another range.