I've got several PIIIs as well, the slowest having two 866HMz CPUs, which will crunch a WU in about 2 days under Linux. Running the "power user app" for Windows discussed in this forum will cut runtimes about 20% even on a PIII, but I must admit that runtimes in the three-digit hours range make you think twice about using those dinosaurs for crunching.
(However, last generation PIIIs should be more energy efficient than early Pentium 4....)
(However, last generation PIIIs should be more energy efficient than early Pentium 4....)
CU
Bikeman
As it happens I am just in the process of retiring a Coppermine 933 MHz host which comes up as "x86 Family 6 Model 8 Stepping 6", so probably a close counsin of Zema88's machine.
The replacement host is a Q9550 "x86 Family 6 Model 23 Stepping 7". I intend to post some notes on the power efficiency considerations. A first taste is that it is well over twenty times more Einstein output for well under twice the electric power.
Something to consider is that the power consumed by the rest of the system beyond the CPU proper must be considered in such comparisons. This tends to work in favor of the faster box, which can amortize that power over more work.
My machine is a Compaq.
Its biggest problem is the 64Mb RAM. Unfortunately I could'n improve that!!!
Do you think PIII are cheaper than modern PC in terms of power consumed?
Something to consider is that the power consumed by the rest of the system beyond the CPU proper must be considered in such comparisons. This tends to work in favor of the faster box, which can amortize that power over more work.
Yes, and multicore CPUs will benefit from this effect even more. But when it comes to comparing P III to P4s, the first P4s were not even faster than the PIIIs they were supposed to replace. So when people are thinking about phasing out their PIIIs, they should do like you did and replace them with modern Core2s (or recent AMD systems). Don't even think about getting a used P4 as replacement :-), at least when you are looking for energy efficiency.
My machine is a Compaq.
Its biggest problem is the 64Mb RAM. Unfortunately I could'n improve that!!!
Do you think PIII are cheaper than modern PC in terms of power consumed?
A lot depends on what else is in the box on both sides of the comparison.
A modern single-core Conroe-class chip with a modest clock rate in a box with a simple motherboard with onboard graphics, and single optical drive, a single hard drive, a decently efficient power supply, and a smallish LCD display would very likely be considerably less electric power than a median PIII box still in use. I organized the purchase of such a machine for an older friend a year ago. Without monitor it cost under $400 from Dell.
But a gamer's PC, with two extravagantly power hungry graphics cards, a quad-core CPU overclocked and overvolted, with an oversized power supply run at too low an output for peak efficiency... can easily burn several times the power of almost any PIII box of that day.
My new box is "pretty lean" except for having a quad-core CPU of higher than minimum clock rate. The box burns a little under 70W at idle, and about 125W running for Einstein jobs. I think it is fairly representative of what someone who wants to build a personal daily use machine but also retire some BOINC-only power wasters might reasonably build. It burns less power than about any three boxes you might reasonably find (even ones running headless), but out-performs almost any three you are likely to find of pre-Conroe vintage.
My particular predecessor Coppermine box burns 75W running Einstein with a user at the keyboard, and 42W at full idle (no work on CPU, graphics card idled because monitor commanded to software power off state). It is not the lowest possible (two hard drives, a somewhat above minimum for the time graphics card) for that era, but plenty from that period would be higher.
My machine is a Compaq.
Its biggest problem is the 64Mb RAM. Unfortunately I could'n improve that!!!
Of course you can :-).
Take out the 64MB stick (or is that 2x32MB) and insert 3x128MB or 2x256MB sticks which you can purchase quite cheaply on ebay for example. You probably have three RAM slots in that Compaq and people basically give away 128MB SDRAM sticks these days. Is it a slot 1 or a socket 370 processor? What is the cpu speed?
Quote:
Do you think PIII are cheaper than modern PC in terms of power consumed
PIIIs are cheaper per unit of time (if you ignore the value of the output) but are far more expensive in terms of cost per unit of work done. Also Coppermine PIIIs are considerably less efficient than Tualatin PIIIs. Unless you are totally unconcerned about the cost of electricity per each task crunched, there is little reason to run a Coppermine PIII - certainly one less than 1.0GHz speed.
As it happens I am just in the process of retiring a Coppermine 933 MHz host which comes up as "x86 Family 6 Model 8 Stepping 6", so probably a close counsin of Zema88's machine.
Your host's benchmarks are 864/1423 whilst those of the Compaq are 289/962. His machine may be a Coppermine 600MHz or maybe even a lower speed Katmai. I hadn't looked at benchmarks previously but having done so now, the Compaq should probably be abandoned ASAP :-).
Quote:
The replacement host is a Q9550 "x86 Family 6 Model 23 Stepping 7". I intend to post some notes on the power efficiency considerations. A first taste is that it is well over twenty times more Einstein output for well under twice the electric power.
I've been quite partial to running some Tualatin PIIIs which I acquired quite cheaply (identical boxes in bulk ~$10 each) around a year ago. They are celeron 1300s and they overclock quite easily under stock conditions to around 1550 - 1600 MHz. Most are running Linux but I just checked one running Windows and its benchmarks were 1522/2578. It averages around 28-29 hours or so per task whilst similar ones running Linux take several hours less, say 24-25 hours or so.
Whilst I've enjoyed tweaking these PIIIs, the time has come to move on so I've been doing a bit of a variation on what you are doing with your Coppermine. I'm not so much pensioning off the PIIIs but rather upgrading them. I have about 60 in production and a number of spares and I've found that I can get a mATX Gigabyte GA-G31M-S2L all-in-one board for about $AUD60 which will slot into the PIII case quite nicely. 2x1G DDR2-800 and a Q6600 complete the deal and I'm using the existing case, PSU, HDD, CDROM, etc. The PSU is not actually the original one (which was only 140W) but is a 250W one which I also acquired as a job lot at a liquidation auction for about $1 each. I bought this particular job lot because each PSU had a 24 pin connector as well as the 4 pin 12V one. These PSUs seem to handle the load without any issues.
So far I've done six conversions. The first one was done over 5 weeks ago and has already accumulated 150K in credit. That agrees pretty much with the current RAC of over 4K. The machine is running Linux at 3.0GHz with stock cooler and close to stock volts. My intention is to pretty much shut down progressively all PIIIs in my fleet (I think I have 2 P4s which will also go) and end up with about 15 Q6600s. I was fortunate to have bought the first six upgrade kits whilst the $AUD was much closer to parity with the $USD than it now is. I'll probably wait for a bit to see if Intel do another price drop any time soon.
Your host's benchmarks are 864/1423 whilst those of the Compaq are 289/962. His machine may be a Coppermine 600MHz or maybe even a lower speed Katmai. I hadn't looked at benchmarks previously but having done so now, the Compaq should probably be abandoned ASAP :-).
That seems odd. I think Family 6 model 8 should mean Coppermine, and I'd be surprised if they sold a seriously low clock rate version by stepping 10. But they certainly did sell Celeron versions, with some intentional crippling. I would just doubt the snapshot of this benchmark (known unreliable on Win98 machines, as both of ours are), save for Zema88's comment about 100 hour task times. Mine is about 44 hours--so his does seem really slow.
But I agree with your basic message. Let me phrase it this way: computers of that vintage are so power inefficient that even if other requirements mean the computer will still be in use 24/7, the extra power required to run BOINC processing on it is a very poor buy.
Quote:
I'm not so much pensioning off the PIIIs but rather upgrading them. I have about 60 in production and a number of spares and I've found that I can get a mATX Gigabyte GA-G31M-S2L all-in-one board for about $AUD60 which will slot into the PIII case quite nicely. 2x1G DDR2-800 and a Q6600 complete the deal and I'm using the existing case, PSU, HDD, CDROM, etc. The PSU is not actually the original one (which was only 140W) but is a 250W one which I also acquired as a job lot at a liquidation auction for about $1 each. I bought this particular job lot because each PSU had a 24 pin connector as well as the 4 pin 12V one. These PSUs seem to handle the load without any issues.
Generally this looks quite good to me. The one point you might want to consider checking is whether your $1.0O power supplies happen to be ones of very low conversion efficiency. If they are, you might actually be money (and carbon...) ahead in not so many months to replace them with an 80+ model appropriately sized to the actual load.
The supply in my build was a Seasonic SS-380GB. At $60 U.S. I was happy to get the promise of reliability and efficiency. I imagine you could find a good enough efficient supply for considerably less. Of course this is a waste of time as well as money if your $1 job lot supplies happen to be highly efficient, but not very long ago many of the supplies out there were more like 60 MINUS than 80 PLUS, and that adds up over time.
For your application the fact that many of them were also unreliable pieces of junk is less of a concern--though for a replacement daily driver--which is my target audience--power supply reliability is a definite consideration (I've had two die, one with an impressive puff of smoke followed by a bright flash, and I only operate about four machines on average over the last ten years.)
As it happens I chose the Q9550 over the popular Q6600 on power, not performance considerations. I confess I had also missed that Q6600 had another price cut (today at my preferred vendor Q9550 is $320 US, vs. $190 for Q6600). For technology reasons I'm pretty sure it has considerably lower standby current. That is utterly irrelevant to a 24/7 BOINC cruncher, but I thought ahead to a possible non-BOINC future in the expected long life of a household machine. I'm less sure what its advantage at a typical operating point is, but suspect it will be enough over the life of the box to amortize the extra initial price. (I'll probably run this household machine for more years than you ought to run your fleet machines, the one it replaces is almost exactly eight years old).
Nit-picking aside, I admire what you are doing, and that you mention it here. About the least good thing that BOINC is doing is inducing the continued operation of many ancient power-inefficient machines which would otherwise retire.
I don't know why, but they told me it was impossible to change the RAM slot...
About the CPU... my benchmarks are:
761 floating point MIPS (Whetstone) per CPU
2060 integer MIPS (Dhrystone) per CPU
What do you say about them?
Measured floating point speed 289.16 million ops/sec
Measured integer speed 961.64 million ops/sec
We were comparing other hosts using that same test. You appear to be citing a different test means, which would need also to be available for comparison hosts to be helpful.
Welcome to
)
Welcome to Einstein@Home!
I've got several PIIIs as well, the slowest having two 866HMz CPUs, which will crunch a WU in about 2 days under Linux. Running the "power user app" for Windows discussed in this forum will cut runtimes about 20% even on a PIII, but I must admit that runtimes in the three-digit hours range make you think twice about using those dinosaurs for crunching.
(However, last generation PIIIs should be more energy efficient than early Pentium 4....)
CU
Bikeman
RE: (However, last
)
As it happens I am just in the process of retiring a Coppermine 933 MHz host which comes up as "x86 Family 6 Model 8 Stepping 6", so probably a close counsin of Zema88's machine.
The replacement host is a Q9550 "x86 Family 6 Model 23 Stepping 7". I intend to post some notes on the power efficiency considerations. A first taste is that it is well over twenty times more Einstein output for well under twice the electric power.
Something to consider is that the power consumed by the rest of the system beyond the CPU proper must be considered in such comparisons. This tends to work in favor of the faster box, which can amortize that power over more work.
My machine is a Compaq. Its
)
My machine is a Compaq.
Its biggest problem is the 64Mb RAM. Unfortunately I could'n improve that!!!
Do you think PIII are cheaper than modern PC in terms of power consumed?
RE: Something to consider
)
Yes, and multicore CPUs will benefit from this effect even more. But when it comes to comparing P III to P4s, the first P4s were not even faster than the PIIIs they were supposed to replace. So when people are thinking about phasing out their PIIIs, they should do like you did and replace them with modern Core2s (or recent AMD systems). Don't even think about getting a used P4 as replacement :-), at least when you are looking for energy efficiency.
CU
Bikeman
RE: My machine is a
)
A lot depends on what else is in the box on both sides of the comparison.
A modern single-core Conroe-class chip with a modest clock rate in a box with a simple motherboard with onboard graphics, and single optical drive, a single hard drive, a decently efficient power supply, and a smallish LCD display would very likely be considerably less electric power than a median PIII box still in use. I organized the purchase of such a machine for an older friend a year ago. Without monitor it cost under $400 from Dell.
But a gamer's PC, with two extravagantly power hungry graphics cards, a quad-core CPU overclocked and overvolted, with an oversized power supply run at too low an output for peak efficiency... can easily burn several times the power of almost any PIII box of that day.
My new box is "pretty lean" except for having a quad-core CPU of higher than minimum clock rate. The box burns a little under 70W at idle, and about 125W running for Einstein jobs. I think it is fairly representative of what someone who wants to build a personal daily use machine but also retire some BOINC-only power wasters might reasonably build. It burns less power than about any three boxes you might reasonably find (even ones running headless), but out-performs almost any three you are likely to find of pre-Conroe vintage.
My particular predecessor Coppermine box burns 75W running Einstein with a user at the keyboard, and 42W at full idle (no work on CPU, graphics card idled because monitor commanded to software power off state). It is not the lowest possible (two hard drives, a somewhat above minimum for the time graphics card) for that era, but plenty from that period would be higher.
RE: My machine is a
)
Of course you can :-).
Take out the 64MB stick (or is that 2x32MB) and insert 3x128MB or 2x256MB sticks which you can purchase quite cheaply on ebay for example. You probably have three RAM slots in that Compaq and people basically give away 128MB SDRAM sticks these days. Is it a slot 1 or a socket 370 processor? What is the cpu speed?
PIIIs are cheaper per unit of time (if you ignore the value of the output) but are far more expensive in terms of cost per unit of work done. Also Coppermine PIIIs are considerably less efficient than Tualatin PIIIs. Unless you are totally unconcerned about the cost of electricity per each task crunched, there is little reason to run a Coppermine PIII - certainly one less than 1.0GHz speed.
Cheers,
Gary.
RE: As it happens I am just
)
Your host's benchmarks are 864/1423 whilst those of the Compaq are 289/962. His machine may be a Coppermine 600MHz or maybe even a lower speed Katmai. I hadn't looked at benchmarks previously but having done so now, the Compaq should probably be abandoned ASAP :-).
I've been quite partial to running some Tualatin PIIIs which I acquired quite cheaply (identical boxes in bulk ~$10 each) around a year ago. They are celeron 1300s and they overclock quite easily under stock conditions to around 1550 - 1600 MHz. Most are running Linux but I just checked one running Windows and its benchmarks were 1522/2578. It averages around 28-29 hours or so per task whilst similar ones running Linux take several hours less, say 24-25 hours or so.
Whilst I've enjoyed tweaking these PIIIs, the time has come to move on so I've been doing a bit of a variation on what you are doing with your Coppermine. I'm not so much pensioning off the PIIIs but rather upgrading them. I have about 60 in production and a number of spares and I've found that I can get a mATX Gigabyte GA-G31M-S2L all-in-one board for about $AUD60 which will slot into the PIII case quite nicely. 2x1G DDR2-800 and a Q6600 complete the deal and I'm using the existing case, PSU, HDD, CDROM, etc. The PSU is not actually the original one (which was only 140W) but is a 250W one which I also acquired as a job lot at a liquidation auction for about $1 each. I bought this particular job lot because each PSU had a 24 pin connector as well as the 4 pin 12V one. These PSUs seem to handle the load without any issues.
So far I've done six conversions. The first one was done over 5 weeks ago and has already accumulated 150K in credit. That agrees pretty much with the current RAC of over 4K. The machine is running Linux at 3.0GHz with stock cooler and close to stock volts. My intention is to pretty much shut down progressively all PIIIs in my fleet (I think I have 2 P4s which will also go) and end up with about 15 Q6600s. I was fortunate to have bought the first six upgrade kits whilst the $AUD was much closer to parity with the $USD than it now is. I'll probably wait for a bit to see if Intel do another price drop any time soon.
Cheers,
Gary.
RE: Your host's benchmarks
)
That seems odd. I think Family 6 model 8 should mean Coppermine, and I'd be surprised if they sold a seriously low clock rate version by stepping 10. But they certainly did sell Celeron versions, with some intentional crippling. I would just doubt the snapshot of this benchmark (known unreliable on Win98 machines, as both of ours are), save for Zema88's comment about 100 hour task times. Mine is about 44 hours--so his does seem really slow.
But I agree with your basic message. Let me phrase it this way: computers of that vintage are so power inefficient that even if other requirements mean the computer will still be in use 24/7, the extra power required to run BOINC processing on it is a very poor buy.
Generally this looks quite good to me. The one point you might want to consider checking is whether your $1.0O power supplies happen to be ones of very low conversion efficiency. If they are, you might actually be money (and carbon...) ahead in not so many months to replace them with an 80+ model appropriately sized to the actual load.
The supply in my build was a Seasonic SS-380GB. At $60 U.S. I was happy to get the promise of reliability and efficiency. I imagine you could find a good enough efficient supply for considerably less. Of course this is a waste of time as well as money if your $1 job lot supplies happen to be highly efficient, but not very long ago many of the supplies out there were more like 60 MINUS than 80 PLUS, and that adds up over time.
For your application the fact that many of them were also unreliable pieces of junk is less of a concern--though for a replacement daily driver--which is my target audience--power supply reliability is a definite consideration (I've had two die, one with an impressive puff of smoke followed by a bright flash, and I only operate about four machines on average over the last ten years.)
As it happens I chose the Q9550 over the popular Q6600 on power, not performance considerations. I confess I had also missed that Q6600 had another price cut (today at my preferred vendor Q9550 is $320 US, vs. $190 for Q6600). For technology reasons I'm pretty sure it has considerably lower standby current. That is utterly irrelevant to a 24/7 BOINC cruncher, but I thought ahead to a possible non-BOINC future in the expected long life of a household machine. I'm less sure what its advantage at a typical operating point is, but suspect it will be enough over the life of the box to amortize the extra initial price. (I'll probably run this household machine for more years than you ought to run your fleet machines, the one it replaces is almost exactly eight years old).
Nit-picking aside, I admire what you are doing, and that you mention it here. About the least good thing that BOINC is doing is inducing the continued operation of many ancient power-inefficient machines which would otherwise retire.
I don't know why, but they
)
I don't know why, but they told me it was impossible to change the RAM slot...
About the CPU... my benchmarks are:
761 floating point MIPS (Whetstone) per CPU
2060 integer MIPS (Dhrystone) per CPU
What do you say about them?
RE: 761 floating point
)
Benchmarked by what tool?
At the moment, Einstein's web page for your host shows:
Measured floating point speed 289.16 million ops/sec
Measured integer speed 961.64 million ops/sec
We were comparing other hosts using that same test. You appear to be citing a different test means, which would need also to be available for comparison hosts to be helpful.