Mark: Those really aren't the systems you should be looking at in a bang for buckish comparison. More enlightening would be the fastest i7-quad core is ranked 30 with 9.7k RAC. The #2 i7 quad is currently my 930 @ 4.1ghz - 8.9k RAC.
The lack of AMD systems in the chart is rather painfully obvious. The 2nd ranked AMD box is a 2xquad core at rank 111 6.6k RAC, the #3 AMD is also a 2xquad core. The 4th highest AMD system is a 6core CPU and is roughly in the same price bucket as an i7 quad. It's ranked at 194 and only has a 5.4k RAC. As the fastest system in its class I assume it's also heavily overclocked and that a direct comparison with the fastest i7 quads is reasonable. What it shows is amd 6 cores being only slightly more than half as fast as i7-quads.
AMD's entry level quads might be more competitive because they fall into a price range where intel's offerings are weak, but for E@H higher performing systems are a no-contest win for Intel.
As a side note, I'll point out that in general benchmarks AMD's chips are much closer so there might be some projects that favor them over Intel.
IIRC all users started with CPU-Crunching, cause there wasn't another option, back in 1999.
(One of my brothers worked on a project for Distributed Computing, GRID Computing and the use of Graphic Processors, 6 years ago, on the Amsterdam (Free) University, lack of funding and lack of knowledge, was the main reason for stopping this project)
First improvement, was made possible by LUNATIC's.
An Optimized SSE2; SSE3; SSSE3x and SSE4.1 application for CPU.
A big accomplishment and a giant step forward in speed.
Then other things like a gaming rig PS2/3, etc. are being used, to crunch data.
BOINC can run on an big range of hardware and their O.S'es
The fact, graphic card's can do a lot off (simple)computations using parallel computing at an incredeble speed, which led to the development of a programming language and the actual use of GPU's.
So which is the percentage of ghost units here? I never had them.
Einstein has got the "Resend lost tasks" function on, which will check between your client_state.xml file (well, actually sched_request*.xml) and the database whether or not all the files claimed to have been assigned to you are actually on your system. If they're not, they'll be resent.
I am using Lunatics applications optimized for SSE3 on my Opteron 1210 running SuSE Linux 11.1 32-bit pae with 5 GB RAM and they are faster than stock apps. I could have installed a 64-bit Linux version but the only project in which 64-bit apps are faster is AQUA.
Tullio
Not all programs can be parallelized. A MonteCarlo simulation cannot be parallelized and I am running 3 projects using MonteCarlo.
Mark: Those really aren't the systems you should be looking at in a bang for buckish comparison. More enlightening would be the fastest i7-quad core is ranked 30 with 9.7k RAC. The #2 i7 quad is currently my 930 @ 4.1ghz - 8.9k RAC.
The lack of AMD systems in the chart is rather painfully obvious. The 2nd ranked AMD box is a 2xquad core at rank 111 6.6k RAC, the #3 AMD is also a 2xquad core. The 4th highest AMD system is a 6core CPU and is roughly in the same price bucket as an i7 quad. It's ranked at 194 and only has a 5.4k RAC. As the fastest system in its class I assume it's also heavily overclocked and that a direct comparison with the fastest i7 quads is reasonable. What it shows is amd 6 cores being only slightly more than half as fast as i7-quads.
AMD's entry level quads might be more competitive because they fall into a price range where intel's offerings are weak, but for E@H higher performing systems are a no-contest win for Intel.
As a side note, I'll point out that in general benchmarks AMD's chips are much closer so there might be some projects that favor them over Intel.
Oh i'm an Intel guy. My 5 main hosts are i7's and I still have few semi-retired core 2 quads lying around.
I realise the top Xeons are the MPI ones and probably cost as much as my entire farm. I was looking to get an i7 980x but they cost as much as two whole i7 920 systems. Its only a chip replacement for my motherboards, but the price is about $1300 (AUD) just for the chip.
A MonteCarlo simulation cannot be parallelized and I am running 3 projects using MonteCarlo.
A Monte Carlo is a series of random steps (under certain constraints), so you can only take one step after the result of the previous one is taken. If each individual step is expensive enough to calculate you can parallelize it (e.g. contains a matrix inversion or other evil stuff). It's probably not a good idea, though, as you want each step to be fast and you always have to pay for the communication overhead.
For this result of random steps to have any meaning you need many complete runs, so you obtain statistically relevant results. And luckily these runs are completely independent of each other, i.e. perfectly parallelizable. In my world "a MonteCarlo simulation" is the sum off all runs, as each one is worthless without the others. That doesn't mean it's easy to do on GPUs, though: it really depends on what has to be calculated during each step. Multiplications and additions are fine, whereas conditional branches are evil ;)
@Mark: if you want to go for a Gulftown you might want to look at the i7 970. It's 133 MHz slower and has a locked multiplier, but shaves off ~120€ of the extreme price. Not suggesting you buy one, though. I certainly wouldn't.
Edit: I've just seen that the price of the i7 950 has been reduced to the former 920 and 930 price point (~250€). That's quite some healthy boost over the 920: 2.67 GHz to 3.06 GHz :)
Yes, 18% should be about the current market share, biased a bit to the high end market due to us selecting only reasonably high performing hosts. But the 10% due to pricing and marketing? Well.. that's a strange way of putting it.
A single Intel 6-core typically gets an RAC of 13 - 14k, whereas the throughput of the best AMD host featuring 2 x 12-cores of RAC ~11k can almost be reached with an Intel 4-core OCed to an estiamted 3.7 GHz (RAC ~10k). So if you want to put it like that: buying an AMD 2 CPU board (340€) and 2 12-cores (2 x 1100€) is more expensive than buying an i7 930 for 250€. So price is the reason we don't see that many top performing AMDs. Oh, and marketing.
MrS
You are not completely correct in his comparisons. Selected AMD machine simply does not work full time, now its RAC fell generally to about 2.5k.
But if you look at the time of execution of individual tasks its ~ 29k sec per one S5GC1 WU. Or ~ 3 WU / day per core. each WU = 250 cr. So it give as ~ 750 RAC per core or ~ 18k RAC if the 24 core machine will operate 24 / 7.
And it's smallish results for AMD (probably this system have some bottleneck not in CPU, may be RAM). For ex my simple and cheap Athlon II X2 gave ~ 1000 RAC per core, when I connect it to the E @ H (~1100-1150 RAC on S5GC1 and ~ 800 RAC on ABP2)
And it's smallish results for AMD (probably this system have some bottleneck not in CPU, may be RAM). For ex my simple and cheap Athlon II X2 gave ~ 1000 RAC per core, when I connect it to the E @ H (~1100-1150 RAC on S5GC1 and ~ 800 RAC on ABP2)
Your Athlon II was probably clocked quite a bit higher than the 12 core monster CPU. These mostly run at 2.3 GHz, which fits the task completion time approximately.
Regarding point 1: you're right, using actual task completion times is much better than RAC. Still the situation doesn't change much that the FP core of the current Athlon architecture is not as strong at E@H as Core 2 (and newer).
Oops, forgot to google specifications of this Opteron processor...
Google says its nominal frequency is only 2.1 GHz. Then it is clear from the difference come, my CPU is not overclocked, but his stock frequency is 3 GHz. The difference in performance per core is just about proportional to the difference in frequency, so that ~ 750 cr/day/core is about normal for him i thinks.
Yep, Core 2(and later) generation of Intel chips still better especially when working in ABP2. For example my old Core 2 Duo T7700 crunch ABP2 WUs an average of 6k seconds each, which gives up to 14.4 WUs per day or up to 1150 RAC per core - the same as my Athlon(if it will works on GW WUs only, try to work on ABP2 is bad idea for this chip - it gives only ~800 cr/day/core), but it operates at a lower frequency (2.4 GHz vs 3 GHz).
Just a situation is not so bad for AMD as one might think from your original post. And given that the AMD chips are cheaper than Intel's (at comparable numbers of cores / frequency), at least here in Russia and in the standard desktop segment of 2-6 cores CPUs (i am not interested in such monsters as the one 12-core processors), the situation is about parity on the performance / price ratio.
That is an average for E @ H. If you watch to a particular job, then ABP2 definitely rules by Intell's chips(or Intel + Cuda). But in GW works AMD may still be the best choice.
Mad_Max: That's only true if your power is free or heavily subsidized. At typical rates a GPUless desktop in the US will use between $150 and $300/year in electricity. In western Europe twice that.
Mark: Those really aren't
)
Mark: Those really aren't the systems you should be looking at in a bang for buckish comparison. More enlightening would be the fastest i7-quad core is ranked 30 with 9.7k RAC. The #2 i7 quad is currently my 930 @ 4.1ghz - 8.9k RAC.
The lack of AMD systems in the chart is rather painfully obvious. The 2nd ranked AMD box is a 2xquad core at rank 111 6.6k RAC, the #3 AMD is also a 2xquad core. The 4th highest AMD system is a 6core CPU and is roughly in the same price bucket as an i7 quad. It's ranked at 194 and only has a 5.4k RAC. As the fastest system in its class I assume it's also heavily overclocked and that a direct comparison with the fastest i7 quads is reasonable. What it shows is amd 6 cores being only slightly more than half as fast as i7-quads.
AMD's entry level quads might be more competitive because they fall into a price range where intel's offerings are weak, but for E@H higher performing systems are a no-contest win for Intel.
As a side note, I'll point out that in general benchmarks AMD's chips are much closer so there might be some projects that favor them over Intel.
IIRC all users started with
)
IIRC all users started with CPU-Crunching, cause there wasn't another option, back in 1999.
(One of my brothers worked on a project for Distributed Computing,
GRID Computing and the use of Graphic Processors, 6 years ago, on the Amsterdam (Free) University, lack of funding and lack of knowledge, was the main reason for stopping this project)
First improvement, was made possible by LUNATIC's.
An Optimized SSE2; SSE3; SSSE3x and SSE4.1 application for CPU.
A big accomplishment and a giant step forward in speed.
Then other things like a gaming rig PS2/3, etc. are being used, to crunch data.
BOINC can run on an big range of hardware and their O.S'es
The fact, graphic card's can do a lot off (simple)computations using parallel computing at an incredeble speed, which led to the development of a programming language and the actual use of GPU's.
RE: So which is the
)
Einstein has got the "Resend lost tasks" function on, which will check between your client_state.xml file (well, actually sched_request*.xml) and the database whether or not all the files claimed to have been assigned to you are actually on your system. If they're not, they'll be resent.
So ghosts here is 0%.
I am using Lunatics
)
I am using Lunatics applications optimized for SSE3 on my Opteron 1210 running SuSE Linux 11.1 32-bit pae with 5 GB RAM and they are faster than stock apps. I could have installed a 64-bit Linux version but the only project in which 64-bit apps are faster is AQUA.
Tullio
Not all programs can be parallelized. A MonteCarlo simulation cannot be parallelized and I am running 3 projects using MonteCarlo.
RE: Mark: Those really
)
Oh i'm an Intel guy. My 5 main hosts are i7's and I still have few semi-retired core 2 quads lying around.
I realise the top Xeons are the MPI ones and probably cost as much as my entire farm. I was looking to get an i7 980x but they cost as much as two whole i7 920 systems. Its only a chip replacement for my motherboards, but the price is about $1300 (AUD) just for the chip.
BOINC blog
RE: A MonteCarlo simulation
)
A Monte Carlo is a series of random steps (under certain constraints), so you can only take one step after the result of the previous one is taken. If each individual step is expensive enough to calculate you can parallelize it (e.g. contains a matrix inversion or other evil stuff). It's probably not a good idea, though, as you want each step to be fast and you always have to pay for the communication overhead.
For this result of random steps to have any meaning you need many complete runs, so you obtain statistically relevant results. And luckily these runs are completely independent of each other, i.e. perfectly parallelizable. In my world "a MonteCarlo simulation" is the sum off all runs, as each one is worthless without the others. That doesn't mean it's easy to do on GPUs, though: it really depends on what has to be calculated during each step. Multiplications and additions are fine, whereas conditional branches are evil ;)
@Mark: if you want to go for a Gulftown you might want to look at the i7 970. It's 133 MHz slower and has a locked multiplier, but shaves off ~120€ of the extreme price. Not suggesting you buy one, though. I certainly wouldn't.
Edit: I've just seen that the price of the i7 950 has been reduced to the former 920 and 930 price point (~250€). That's quite some healthy boost over the 920: 2.67 GHz to 3.06 GHz :)
MrS
Scanning for our furry friends since Jan 2002
RE: Yes, 18% should be
)
You are not completely correct in his comparisons. Selected AMD machine simply does not work full time, now its RAC fell generally to about 2.5k.
But if you look at the time of execution of individual tasks its ~ 29k sec per one S5GC1 WU. Or ~ 3 WU / day per core. each WU = 250 cr. So it give as ~ 750 RAC per core or ~ 18k RAC if the 24 core machine will operate 24 / 7.
And it's smallish results for AMD (probably this system have some bottleneck not in CPU, may be RAM). For ex my simple and cheap Athlon II X2 gave ~ 1000 RAC per core, when I connect it to the E @ H (~1100-1150 RAC on S5GC1 and ~ 800 RAC on ABP2)
RE: And it's smallish
)
Your Athlon II was probably clocked quite a bit higher than the 12 core monster CPU. These mostly run at 2.3 GHz, which fits the task completion time approximately.
Regarding point 1: you're right, using actual task completion times is much better than RAC. Still the situation doesn't change much that the FP core of the current Athlon architecture is not as strong at E@H as Core 2 (and newer).
MrS
Scanning for our furry friends since Jan 2002
Oops, forgot to google
)
Oops, forgot to google specifications of this Opteron processor...
Google says its nominal frequency is only 2.1 GHz. Then it is clear from the difference come, my CPU is not overclocked, but his stock frequency is 3 GHz. The difference in performance per core is just about proportional to the difference in frequency, so that ~ 750 cr/day/core is about normal for him i thinks.
Yep, Core 2(and later) generation of Intel chips still better especially when working in ABP2. For example my old Core 2 Duo T7700 crunch ABP2 WUs an average of 6k seconds each, which gives up to 14.4 WUs per day or up to 1150 RAC per core - the same as my Athlon(if it will works on GW WUs only, try to work on ABP2 is bad idea for this chip - it gives only ~800 cr/day/core), but it operates at a lower frequency (2.4 GHz vs 3 GHz).
Just a situation is not so bad for AMD as one might think from your original post. And given that the AMD chips are cheaper than Intel's (at comparable numbers of cores / frequency), at least here in Russia and in the standard desktop segment of 2-6 cores CPUs (i am not interested in such monsters as the one 12-core processors), the situation is about parity on the performance / price ratio.
That is an average for E @ H. If you watch to a particular job, then ABP2 definitely rules by Intell's chips(or Intel + Cuda). But in GW works AMD may still be the best choice.
Mad_Max: That's only true if
)
Mad_Max: That's only true if your power is free or heavily subsidized. At typical rates a GPUless desktop in the US will use between $150 and $300/year in electricity. In western Europe twice that.