Do dual CPU machines use all their processing power when running Einstein@Home?
10 Jan 2006 3:20:58 UTC
Topic 190572
(moderation:
I was looking at new computers today and was surprised at the number of new models with dual processors like the AMD 64*2.
My question is do DC programs like Seti and Einstein take full advantage of a computer having two processors? In other words, would I get more WUs done in a year with a computer that has one 3.2Ghz processor or one that has two 2.5 Ghz processors?
Do dual CPU machines use all their processing power when running
)
Two. There is a preference setting for allowing BOINC to use however many CPUs you have in one host. There is at least one SETI cruncher who has a 64-CPU machine.
You get the _most_ benefit from "true" Dual-Core CPUs such as the AMD X2 3800 or 4400; the Hyperthreading in some Intel P4's only gives you somewhere around 5-20% boost, not 100%. Someone on the SETI boards just went through a lot of questions recently on which CPU to pick, the consensus was that in dual-cores, the AMD 3800 is the best "bang for the buck", the 4200 was pointless, the 4400 was best if you weren't going to overclock and could afford the price difference. The Intel _true_ dual-cores are still too expensive, although there's a lot to be said for their server-level chips at the high end...
RE: My question is do DC
)
A P4 3.2 GHz with HT will do about the same amount of work, maybe a little less than a 3.5 GHz system.
Because of the variances in the granting of credit it is real hard to get a really good fix on the numbers. There is also the complication on the fact that the AMD and Intel chips do slightly better on one project or another ...
If you really want to rip through the work at Einstein@Home get a G5 ... :)
When I can afford it again, I will likely be upgrading all my "home built" systems with dual core processors in the 3.2/3200+ neighgorhood to take advantage of the ability to run each result wihin its own dedicated CPU. Even if I just go P4 3.2 GHz HT to dual core P4 3.2 GHz I can probably expect a 30% increase in throughput overall.
So, it all boils down to what you can afford ...
RE: Two. There is a
)
What level of processor? The sticker price on that sort of setup prohibitively high for a personal purchase while new, and most institutions that did buy one would presumably keep it fairly busy Id imagine.
RE: What level of
)
Not really. Prices start at around 150€. This is for an AMD Athlon 64 3000+, see this
shop.
edit: This thread was about dual core processors. They start at 330€. Same link. Look for "X2".
edit: D'oh. Read, then post! I read "64 bit CPU".
Michael
Team Linux Users Everywhere
RE: I was looking at new
)
Have a look at http://einstein.phys.uwm.edu/top_hosts.php
and see, this are all multiprocessors or multicores. For myself I'm running an AMD Dualcore http://einsteinathome.org/host/411398 and both cores are doing WUs.
hth faeshn
Do you like to join a team? http://einstein.phys.uwm.edu/team_display.php?teamid=524
RE: RE: What level of
)
The cheapest x86_64 processor is a Sempron 64 for soket 754, starting at 65 € http://www.geizhals.at/eu/?cat=cpuamda64&sort=p :-)
And the cheapest Dual Core processor is the new Intel Yonah, but maybe a mainboard for it will be more expensive.
Intel® Core™ Duo processor
Mobile (FCBGA6 / FCPGA6)
T2600 (2M L2 cache 2.16 GHz 667 MHz FSB 65nm) - $637 -
T2500 (2M L2 cache 2 GHz 667 MHz FSB 65nm) - $423 -
T2400 (2M L2 cache 1.83 GHz 667 MHz FSB 65nm) - $294 -
T2300 (2M L2 cache 1.66 GHz 667 MHz FSB 65nm) - $241 -
http://www.intel.com/intel/finance/pricelist/
Hmmm. I'm assuming those
)
Hmmm. I'm assuming those chips are based on the same P3 derived architecture as the P-M. How well do they perform, I know the old p3 was roughly 1:1 equivilant with an athlon, but don't know much about the new ones.
RE: the Hyperthreading in
)
As was mentioned, this depends rather drastically on the problem. For SETI@home, a large-cache HT P4 such as the 3.2 GHz Extreme Edition (2Megabyte cache) I run gets a really dramatic benefit from running "2 processor" HT vs. HT disabled--well 20%. Yet more dramatic is the apparent synergy benefit of running HT on this machine when one job is SETI and the other Einstein. The Einstein side seems to run the same speed as one side would run were there two Einsteins running, but the SETI side speeds up substantially (must mean that Einstein and SETI use of the indepently scheduled execution resources differs enough to get some useful overlap).
For my machine, SETI typical times when running SETI on both sides in HT appear to cluster around 65 minutes, while the HT SETI time when running Einstein on the other side looks more like 48 minutes.
I think a lot depends on whether one greatly alters the cache hit rate between HT and non-HT. If the hit rate was pretty high as a uniprocessor, but drops drastically in HT, that can gobble up the HT benefit in a hurry. I suspect that a 512K or 1M cache P4 running HT on SETI thus gets far less HT benefit thatn the 2M cache models. As I doubt the Einstein locality of reference is the same, I suspect any SETI results are quite wide of the mark in predicting Einstein HT behavior.
RE: I was looking at new
)
I've just placed a purchase order for a new LIGO data analysis cluster at UWM. It will be quantity 640 AMD Opteron 175 systems: a total of 1280 CPU cores. The systems are based on Supermicro H8SSL-i motherboards with 2 x 1GB of DDR/400 unregistered ECC memory and a single 80GB SATA drive.
(By the way, the Opteron 175 is identical silicon to the AMD64 X2 4400+.)
Director, Einstein@Home
RE: RE: I was looking at
)
Any detail on software?
Tullio