[XPM] [ANN] Primecoin High Performance | HP14 released!

[ReCat]

Well, I mined for 4 days straight using yPool's Win64 miner on my rack server, and all that it rewarded me is 0.2 XPM. :I

I could have mined a block in this time!

Fuck!

[mikaelh]

Mikael: in the below code block should it be nLayerSeq < nBiTwinCC2Layers or nLayerSeq < nBiTwinCC1Layers? It looks like it should be a 1, and rdebourbon made this change in his miner, but it seems the chains/d drop when I do this.. so curious if you knew what was going on there. The double check in the first if statement seems redundant too.

Code:

                if (nLayerSeq < nBiTwinCC1Layers && nLayerSeq < nBiTwinCC2Layers)
                {
                    for (unsigned int nWord = nMinWord; nWord < nMaxWord; nWord++)
                    {
                        vfCompositeCunningham1[nWord] |= vfCompositeLayerCC1[nWord];
                        vfCompositeCunningham2[nWord] |= vfCompositeLayerCC2[nWord];
                        vfCompositeBiTwin[nWord] |= vfCompositeLayerCC1[nWord] | vfCompositeLayerCC2[nWord];
                    }
                }
                else if (nLayerSeq < nBiTwinCC2Layers)

Well, the second if-statement should say nLayerSeq < nBiTwinCC1Layers. I read rdebourbon's release notes and saw the bug in the code earlier already.

And yes, the first if-statement could be simplified to nLayerSeq < nBiTwinCC2Layers since that implies nLayerSeq < nBiTwinCC1Layers, if that's what you meant.

The bug causes the sieve to produce BiTwin candidates for 8-chains instead of 9-chains on mainnet. That means there will be more TWN candidates but they are a lot less likely to produce blocks. More candidates means that more primality tests are done which in turn falsely inflates the chains/day estimate. Basically this is the big bug in HP10 that I've been looking for. Once the bug is fixed, the miner will be producing more blocks and chains/day will be a lot closer to reality.

[cabin]

Mikael: in the below code block should it be nLayerSeq < nBiTwinCC2Layers or nLayerSeq < nBiTwinCC1Layers? It looks like it should be a 1, and rdebourbon made this change in his miner, but it seems the chains/d drop when I do this.. so curious if you knew what was going on there. The double check in the first if statement seems redundant too.

Code:

                if (nLayerSeq < nBiTwinCC1Layers && nLayerSeq < nBiTwinCC2Layers)
                {
                    for (unsigned int nWord = nMinWord; nWord < nMaxWord; nWord++)
                    {
                        vfCompositeCunningham1[nWord] |= vfCompositeLayerCC1[nWord];
                        vfCompositeCunningham2[nWord] |= vfCompositeLayerCC2[nWord];
                        vfCompositeBiTwin[nWord] |= vfCompositeLayerCC1[nWord] | vfCompositeLayerCC2[nWord];
                    }
                }
                else if (nLayerSeq < nBiTwinCC2Layers)

Well, the second if-statement should say nLayerSeq < nBiTwinCC1Layers. I read rdebourbon's release notes and saw the bug in the code earlier already.

And yes, the first if-statement could be simplified to nLayerSeq < nBiTwinCC2Layers since that implies nLayerSeq < nBiTwinCC1Layers, if that's what you meant.

The bug causes the sieve to produce BiTwin candidates for 8-chains instead of 9-chains on mainnet. That means there will be more TWN candidates but they are a lot less likely to produce blocks. More candidates means that more primality tests are done which in turn falsely inflates the chains/day estimate. Basically this is the big bug in HP10 that I've been looking for. Once the bug is fixed, the miner will be producing more blocks and chains/day will be a lot closer to reality.

Thanks, that clears things up!

[mikaelh]

Well, I pushed the fix for the big bug to github. I'm also changing the default value of 'sieveextensions' to 9. That seems to be the new sweet spot now that the bug in the sieve is fixed. If all goes well, I will probably do a release some time tomorrow.

[Ifaistos]

pulled the code from github and build it
cpd wend from 0.30 to 0.19 no idea about block finds though, as i started yesterday with H10 on 5 similar machines (old dual xeon servers)

{
    "blocks" : 163249,
    "chainspermin" : 1,
    "chainsperday" : 0.19348673,
    "currentblocksize" : 1000,
    "currentblocktx" : 0,
    "difficulty" : 9.86140740,
    "errors" : "",
    "generate" : true,
    "genproclimit" : -1,
    "primespersec" : 380,
    "pooledtx" : 0,
    "sieveextensions" : 9,
    "sievepercentage" : 10,
    "sievesize" : 1000000,
    "testnet" : false
}

[Ifaistos]

I've spend the day yesterday going over this thread (all 113 pages of it  ) trying to figure out optimal settings.
(People mention several times in the thread that the defaults are pretty much the best settings)
I have a number of old dual Xeon 3.4Ghz servers with 16k L1 and 1MB L2 and 2GB ram and i am not sure if the defaults are actually the best option for them, as newer cpu's have much bigger L1 and L2 caches.

As far as i understood without going over the code, sievesize is the pool of primes we look at, sievepercentage is the % of the numbers we look from the sieve for candidates for the chains.
I am not sure what the sieveextensions parameter means (there was a mention but can't locate it)
What is not clear to me and i would be great full if someone could explain, is how sievesize translates to actual memory footprint ?
Cause i believe that if that does not fit in the L2 cache things would slow down considerable (then again i might be wrong)

Thanks for your time

[Trillium]

I've spend the day yesterday going over this thread (all 113 pages of it   ) trying to figure out optimal settings.
(People mention several times in the thread that the defaults are pretty much the best settings)
I have a number of old dual Xeon 3.4Ghz servers with 16k L1 and 1MB L2 and 2GB ram and i am not sure if the defaults are actually the best option for them, as newer cpu's have much bigger L1 and L2 caches.

As far as i understood without going over the code, sievesize is the pool of primes we look at, sievepercentage is the % of the numbers we look from the sieve for candidates for the chains.
I am not sure what the sieveextensions parameter means (there was a mention but can't locate it)
What is not clear to me and i would be great full if someone could explain, is how sievesize translates to actual memory footprint ?
Cause i believe that if that does not fit in the L2 cache things would slow down considerable (then again i might be wrong)

Thanks for your time

While I can't answer your question re cache optimization, I can tell you that those CPU's wont break even against electricity costs unless you have almost-free or free electricity. I have a pile of 10 rackmount servers behind me with those kinds of CPU (release dates roughly 2003 - 2007) and they're simply not worth mining on. An i7 4770k is about as fast of ten of those older single core xeons and only uses <120 watts even before you overclock it.

My work laptop (i5-560M) gets 0.8 chainsperday and uses 40 watts from the wall, it's faster than about 500 watts of my old rackmount servers (not to mention essentially silent and 1/100th the size).

[peoplecoin]

I heard that someone sell new block hashes?

[Carra23]

I have been continuously mining from day 1. I have never optimized anything, and just run the qt with setgenerate true. Is there anything I can do to improve?

[Ifaistos]

While I can't answer your question re cache optimization, I can tell you that those CPU's wont break even against electricity costs unless you have almost-free or free electricity. I have a pile of 10 rackmount servers behind me with those kinds of CPU (release dates roughly 2003 - 2007) and they're simply not worth mining on. An i7 4770k is about as fast of ten of those older single core xeons and only uses <120 watts even before you overclock it.

My work laptop (i5-560M) gets 0.8 chainsperday and uses 40 watts from the wall, it's faster than about 500 watts of my old rackmount servers (not to mention essentially silent and 1/100th the size).

You are right that it does not make sense without free electricity given current difficulty and price ~0.30$

I measured them with a Voltcratf energy logger at around 300W with both cpus and hyperthreading on, running H10
With the 'old' H10 they do around 0.3 cpd, with the version from github goes down to 0.22
So doing the math gives us

block/day = (1 - 0.86185294 ) * 0.22924828 = 0.031669976 which is around 1 block every 30 days
 12 xpm * 0.30$ =3.6$/month

power consumption

 330W = 7.92 Kw/h a day = 237.6 Kw/h per month

so to *break even* (no profit) you need an electricity price of less than  0.015 Kw/h (237.6/3.6) and without factoring in things like cooling,internet connection etc.

If on the other hand if you don't have to pay for any of the running costs with 10 machines and current prices you could get around 36$/month  from some piece of junk equipment lol  

[mikaelh]

-hp11 is out! Download links are on the first page as usual.

Changes in -hp11:
 * Fixed a bug in the BiTwin filter that was inflating the chains/day estimate and reducing the chances of finding a block
 * Slightly improved the accuracy of the prime probability estimate with regards to the 'sieveextensions' parameter
 * Other small fixes and improvements from cabin

This is mainly a bugfix release without any new features. The chains/day number does go down with this release now that the bug in HP10's sieve has been fixed. Nevertheless, the actual chance of finding a block is improved. Furthermore, the reduced chains/day estimate is actually a lot closer to reality now.

[Trillium]

-hp11 is out! Download links are on the first page as usual.

Changes in -hp11:
 * Fixed a bug in the BiTwin filter that was inflating the chains/day estimate and reducing the chances of finding a block
 * Slightly improved the accuracy of the prime probability estimate with regards to the 'sieveextensions' parameter
 * Other small fixes and improvements from cabin

This is mainly a bugfix release without any new features. The chains/day number does go down with this release now that the bug in HP10's sieve has been fixed. Nevertheless, the actual chance of finding a block is improved. Furthermore, the reduced chains/day estimate is actually a lot closer to reality now.

Our god has spoken. Bow down and receive His gift graciously.

[arnuschky]

-hp11 is out! Download links are on the first page as usual.

Changes in -hp11:
 * Fixed a bug in the BiTwin filter that was inflating the chains/day estimate and reducing the chances of finding a block
 * Slightly improved the accuracy of the prime probability estimate with regards to the 'sieveextensions' parameter
 * Other small fixes and improvements from cabin

This is mainly a bugfix release without any new features. The chains/day number does go down with this release now that the bug in HP10's sieve has been fixed. Nevertheless, the actual chance of finding a block is improved. Furthermore, the reduced chains/day estimate is actually a lot closer to reality now.

Thank you. Not only for making this, but also for continuing to support Primecoin, the hp client, and all the other work!

[hasle2]

I have been continuously mining from day 1. I have never optimized anything, and just run the qt with setgenerate true. Is there anything I can do to improve?

As long as you are always running the latest hp then you are pretty much doing the best you can. The default parameters are pretty much optimal for most processors.

[ReCat]

Well, No wonder my rack server was mining so many blocks! Although it's from 2006 the CPU beat even the fastest i7 at prime numbers.

Interesting stuff!

[Trillium]

https://i.imgur.com/3GDcG0N.png

Well, No wonder my rack server was mining so many blocks! Although it's from 2006 the CPU beat even the fastest i7 at prime numbers.

Interesting stuff!

I think for primecoin integer math is also important. IIRC mikaelh has mentioned this before (specifically in regards to integer / floating point ops). I don't think that cpumark prime score translates directly to primecoin performance. How many watts on your rack? In my experience servers are not very efficient but if I recall correctly your server (from previous pics) was mostly just motherboard+CPUs so perhaps its more efficient than most systems from the big vendors (HP, IBM, Dell etc). I see a lot of rackmount servers that use 200+ watts of juice while idling and I wonder where exactly that power is going... Oh and keep in mind that a 3770k can overclock 25-40%+ without much effort or additional cooling. I think that cpumark disregards submitted results where an overclock is present to prevent skewing of the results.

[ReCat]

https://i.imgur.com/3GDcG0N.png

Well, No wonder my rack server was mining so many blocks! Although it's from 2006 the CPU beat even the fastest i7 at prime numbers.

Interesting stuff!

I think for primecoin integer math is also important. IIRC mikaelh has mentioned this before (specifically in regards to integer / floating point ops). I don't think that cpumark prime score translates directly to primecoin performance. How many watts on your rack? In my experience servers are not very efficient but if I recall correctly your server (from previous pics) was mostly just motherboard+CPUs so perhaps its more efficient than most systems from the big vendors (HP, IBM, Dell etc). I see a lot of rackmount servers that use 200+ watts of juice while idling and I wonder where exactly that power is going... Oh and keep in mind that a 3770k can overclock 25-40%+ without much effort or additional cooling. I think that cpumark disregards submitted results where an overclock is present to prevent skewing of the results.

It's a Dual-Processor Motherboard. Each CPU pulls a maximum of 50 watts. About 220 watts are drawn when on full load, 170 otherwise. I measured it with a Killawatt meter on site.

You can't really beat that when considering I got the entire server for just 150 bucks.

[FoldingTime]

Well one month of continuous solo mining on the latest HP* and found 8 blocks on Xeon E3-1230. Also there appears to be no continuity on when I'll find a block:

Sun Aug 11 22:11:51 CEST 2013
Mon Aug 12 23:37:51 CEST 2013
Sun Aug 18 11:19:38 CEST 2013
Tue Aug 27 17:31:03 CEST 2013
Wed Aug 28 11:53:26 CEST 2013
Fri Sep  6 08:24:34 CEST 2013
Sun Sep 15 08:11:59 CEST 2013
Sun Sep 15 10:23:27 CEST 2013

[Ifaistos]

What's the CPD you are getting ?
Want to see if the formula used to "predict" blocks has actually any meaning, although variance can be a bitch

[Ifaistos]

https://i.imgur.com/3GDcG0N.png

Well, No wonder my rack server was mining so many blocks! Although it's from 2006 the CPU beat even the fastest i7 at prime numbers.

Interesting stuff!

I think for primecoin integer math is also important. IIRC mikaelh has mentioned this before (specifically in regards to integer / floating point ops). I don't think that cpumark prime score translates directly to primecoin performance. How many watts on your rack? In my experience servers are not very efficient but if I recall correctly your server (from previous pics) was mostly just motherboard+CPUs so perhaps its more efficient than most systems from the big vendors (HP, IBM, Dell etc). I see a lot of rackmount servers that use 200+ watts of juice while idling and I wonder where exactly that power is going... Oh and keep in mind that a 3770k can overclock 25-40%+ without much effort or additional cooling. I think that cpumark disregards submitted results where an overclock is present to prevent skewing of the results.

It's a Dual-Processor Motherboard. Each CPU pulls a maximum of 50 watts. About 220 watts are drawn when on full load, 170 otherwise. I measured it with a Killawatt meter on site.

You can't really beat that when considering I got the entire server for just 150 bucks.

I am getting similar numbers in terms of power (250W) from a Dell poweredge 2900 with 2  E5405  @ 2.00GHz.
HP11 gives a 1.5 in CPD but no block yet after a day and  half of running