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

[pocesar]

Does the version hp9 takes advantage of AVX on core i processors? if not, hp10 will include those?

Any software can use AVX if you compile it yourself with a decent compiler.

In a precompiled release, it's possible to have separate code paths with such optimizations, but I think such releases are unlikely at this point. AVX is just one technological improvement among others, like the numerous SSE iterations, and it would take dozens of different binaries to cover all possible optimizations. Also, besides the instruction set, there are optimizations for scheduling and cache sizes. Speaking of which, Primecoin performance is apparently sensitive to L1 cache; there are options like sievesize that may be optimized for a given cache size without recompiling.

In general, though, if you want something optimized for your particular machine, you need to compile it yourself.

alright, I will try some switches, thanks. just asked because there are some ASM includes that are specific to AVX, and they perform really really well and the opcode can be easily identified (and skipped, in code if not found) and used. scrypt and sha256 already got those optimized versions.

[1713427450]

1713427450

[1713427450]

1713427450

[1713427450]

1713427450

[1713427450]

1713427450

[Sharky444]

Is there a parameter to disable the default debug mode? I think it costs performance (a lot of writes every second to debug.log).

[Trillium]

Is there a parameter to disable the default debug mode? I think it costs performance (a lot of writes every second to debug.log).

Appending text to the debug log uses nil system resources. Don't worry about it.

[Tamis]

I remember people having performance problems on machines with a lot of cores.
I'm tempted to get a Dual Intel Xeon L5420, that represents 8 dedicated and 16 virtual cores.

Would I have performance problems with this setup ?

[gigawatt]

I'm not sure if I did this right, but I made a tweak to the "primemeter" section of the code to get a peek at what solving difficulty+1 would be like.  If this is right, all I have to say is "oof".
If I've done this wrong, please tell me.
I really really hope I've done this wrong.

Original main.cpp:

Code:

if (nMillisNow - nLogTime > 59000)
{
    nLogTime = nMillisNow;
    printf("%s primemeter %9.0f prime/h %9.0f test/h %4.0f %d-chains/h %3.6f chain/d\n", DateTimeStrFormat("%Y-%m-%d %H:%M:%S", nLogTime / 1000).c_str(), dPrimesPerMinute * 60.0, dTestsPerMinute * 60.0, dChainsPerMinute * 60.0, nStatsChainLength, dChainsPerDay);
}

Tweaked main.cpp:

Code:

if (nMillisNow - nLogTime > 59000)
{
    nLogTime = nMillisNow;
    double next_prob = EstimateCandidatePrimeProbability(nPrimorialMultiplier, TargetGetLength(pblock->nBits));
    double est_next = dChainsPerDay * next_prob;
    printf("%s primemeter %9.0f prime/h %9.0f test/h %4.0f %d-chains/h %3.6f chain/d %u RoundPrimorial %3.6f est_next %3.6f next_prob\n", DateTimeStrFormat("%Y-%m-%d %H:%M:%S", nLogTime / 1000).c_str(), dPrimesPerMinute * 60.0, dTestsPerMinute * 60.0, dChainsPerMinute * 60.0, nStatsChainLength, dChainsPerDay, nPrimorialMultiplier, est_next, next_prob);
}

Result (1 vCPU VPS):

Code:

2013-08-22 15:23:44 primemeter    892352 prime/h  14610791 test/h   60 5-chains/h 0.153438 chain/d 47 RoundPrimorial 0.013593 est_next 0.088589 next_prob
2013-08-22 15:24:44 primemeter    878830 prime/h  14379729 test/h    0 5-chains/h 0.152076 chain/d 41 RoundPrimorial 0.013928 est_next 0.091584 next_prob
2013-08-22 15:25:44 primemeter    885159 prime/h  14521363 test/h  180 5-chains/h 0.152199 chain/d 43 RoundPrimorial 0.013710 est_next 0.090079 next_prob
2013-08-22 15:26:44 primemeter    877952 prime/h  14345645 test/h   60 5-chains/h 0.150436 chain/d 43 RoundPrimorial 0.013551 est_next 0.090079 next_prob
2013-08-22 15:27:44 primemeter    879141 prime/h  14404795 test/h   60 5-chains/h 0.152234 chain/d 43 RoundPrimorial 0.013713 est_next 0.090079 next_prob

So yeah, 0.152 chain/d now at diff 9, only 0.013 chain/d at diff 10?
I really hope I'm wrong.


edit: I did spot an inconsistency issue.  chain/d is measured over the entire 60 second interval of various values for RoundPrimorial, but the est_next/next_prob values are based on the current value of it.  I imaging it can skew the results +/- 10%.

[crendore]

I remember people having performance problems on machines with a lot of cores.
I'm tempted to get a Dual Intel Xeon L5420, that represents 8 dedicated and 16 virtual cores.

Would I have performance problems with this setup ?

If you're talking about buying a new computer, i wouldn't bother with Xeon, just go with an i7 or something, much cheaper.

[Tamis]

I remember people having performance problems on machines with a lot of cores.
I'm tempted to get a Dual Intel Xeon L5420, that represents 8 dedicated and 16 virtual cores.

Would I have performance problems with this setup ?

If you're talking about buying a new computer, i wouldn't bother with Xeon, just go with an i7 or something, much cheaper.

No just getting a dedicated server to play with. Already got an i7 3770k at home.

[Stinky_Pete]

I remember people having performance problems on machines with a lot of cores.
I'm tempted to get a Dual Intel Xeon L5420, that represents 8 dedicated and 16 virtual cores.

Would I have performance problems with this setup ?

I've not had problems, currently getting about 1 block per day.

[ReCat]

I remember people having performance problems on machines with a lot of cores.
I'm tempted to get a Dual Intel Xeon L5420, that represents 8 dedicated and 16 virtual cores.

Would I have performance problems with this setup ?

I mined primecoin for a while on the L5420. It mines perfectly fine. Was slightly faster than my friend's Sandy Bridge Core i5. It's great considering rack servers with the L5420 cost less than 150 USD total.

[Tamis]

I remember people having performance problems on machines with a lot of cores.
I'm tempted to get a Dual Intel Xeon L5420, that represents 8 dedicated and 16 virtual cores.

Would I have performance problems with this setup ?

I've not had problems, currently getting about 1 block per day.

Thanks Stinky !

As a rule of thumb what is the ram requirement per core ?

[shazbits]

I remember people having performance problems on machines with a lot of cores.
I'm tempted to get a Dual Intel Xeon L5420, that represents 8 dedicated and 16 virtual cores.

Would I have performance problems with this setup ?

I mined primecoin for a while on the L5420. It mines perfectly fine. Was slightly faster than my friend's Sandy Bridge Core i5. It's great considering rack servers with the L5420 cost less than 150 USD total.

Where do you buy a rack server with a Xeon L5420 for $150? Used one on ebay? How's the power consumption at the wall?

I've not had problems, currently getting about 1 block per day.

1 block per day with a single L5420? Somehow I doubt that because I was getting 1 block every few days a couple weeks back, using an E5-1650.

[Stinky_Pete]

Dual Intel Xeon

[atariguy]

I remember people having performance problems on machines with a lot of cores.
I'm tempted to get a Dual Intel Xeon L5420, that represents 8 dedicated and 16 virtual cores.

Would I have performance problems with this setup ?

I've not had problems, currently getting about 1 block per day.

Thanks Stinky !

As a rule of thumb what is the ram requirement per core ?

The most that the Primecoin-Qt process uses with 8 cores on any of the machines I'm running is 112 MB.

[gigawatt]

As a rule of thumb what is the ram requirement per core ?

The most that the Primecoin-Qt process uses with 8 cores on any of the machines I'm running is 112 MB.

Running it is cake.  Compiling it will require at least 1 GB of memory (RAM/Swap).
You can compile it with less (for example, on an OpenVZ VPS), but it'll take forever.

[Tamis]

Thank you for your answer you two !

Did you notice any difference between the already compiled primecoind and the one compiled on the server ?
Some vps I have only have 256mb of ram and swap so I can't compile on them so I use the already compiled one and did not notice any difference in performance.

Why compile on an openvz ? You will get suspended for unfair cpu usage on those !

[gigawatt]

Thank you for your answer you two !

Did you notice any difference between the already compiled primecoind and the one compiled on the server ?
Some vps I have only have 256mb of ram and swap so I can't compile on them so I use the already compiled one and did not notice any difference in performance.

Why compile on an openvz ? You will get suspended for unfair cpu usage on those !

Usually compiled locally gives better results if you include compile flags like "-mtune=native -march=native" so the binary is tailored to your hardware.

The reason it's difficult to compile on OpenVZ is because you cannot use swapon to use a swapfile, you're forced to use only the memory you have available.  You can compile with certain memory restrictions, but it makes the compile literally take over 20 times as long.  It ends up triggering a garbage collection at almost every step of the process and limits how much working space it's allowed.  The compiled binary will work perfectly fine, but the compile time goes through the roof.

Also, I don't think the virtualization platform matters as much for account suspension as much as the network administration does.  Some virtualizations may have more/less support for load monitoring, but it eventually all comes down to whoever is running the show.

[ivanlabrie]

I think I'm gonna jump on those used dual xeon servers too...Tamis, they are quite good, I know a guy who has one.
Expect a bit better performance vs your 3770k.
Not half bad for 300usd at ebay.

[Entz]

If you are going the used/ebay route, At $300 for those I would expect the newly off lease HP G180 G6 dual L5639 servers @399  to be a better buy (12C/24T 1 gen newer, lower power etc)

Edit Looked: 300 is high, the dual 5420s should be under $200  (i.e. a bunch of SE1101's for 150). That is a good price if you have a spot for them

[ivanlabrie]

Thanks man! Yeah, found a seller...399 is quite sweet for those dual hexa core rigs.

[Tamis]

I think I'm gonna jump on those used dual xeon servers too...Tamis, they are quite good, I know a guy who has one.
Expect a bit better performance vs your 3770k.
Not half bad for 300usd at ebay.

Not bad indeed, but I live in France so the shipping is a killer :/

My 3770k just found two blocks in 7 hours   To think what I was getting on ypool when I tried it for a few (lost) hours... just dust on this pool !