I've reduced about 75% of the CPU utilisation, and GPU tasks sped up. With only 2 out of 8 physical CPU cores utilised, my 8800GT crunch thru a CUDA task on an estimated sub-8,000 sec, down from the usual 11,000-sec GPU time. Pretty consistent to Apple's underclocked nVidia 8800GT firmware. So what happened here...? CPU contention problem...?
I'd guess the problem is memory contention. Eight physical cores run E@H tasks can consume a lot of memory bandwidth.
I've reduced about 75% of the CPU utilisation, and GPU tasks sped up. With only 2 out of 8 physical CPU cores utilised, my 8800GT crunch thru a CUDA task on an estimated sub-8,000 sec, down from the usual 11,000-sec GPU time. Pretty consistent to Apple's underclocked nVidia 8800GT firmware. So what happened here...? CPU contention problem...?
I'd guess the problem is memory contention. Eight physical cores run E@H tasks can consume a lot of memory bandwidth.
CU
HBE
But Nehalem processors have largely addressed the memory contention problem, haven't they...? I used to have a Clovertown-based Mac Pro, that was a real memory contention problem.
For the most part they have. The 2 channel controller on an quadcore LGA1156 chip can handle 75% of the peak theoretical memory demands of the CPU at DDR3-1333; and that outside of benchmarking it was almost impossible to reach levels high enough to saturate the memory bus. (Intel didn't say exactly which of the chips they were referring to to annanadtech.) The limited bandwidth from the dual channel controller is why they will only be releasing hexacore processors on LGA1366 (Presumably with an uptick in the officially supported ram speed to something above DDR3-1066.)
But Nehalem processors have largely addressed the memory contention problem, haven't they...? I used to have a Clovertown-based Mac Pro, that was a real memory contention problem.
Ah ok, but if we are talking Nehalem CPUs, what about hyperthreading? Did the first, slower, CUDA task runtime correspond to a setup with hyperthreading enabled?
Ah ok, but if we are talking Nehalem CPUs, what about hyperthreading? Did the first, slower, CUDA task runtime correspond to a setup with hyperthreading enabled?
CU
HBE
This one I wouldn't be sure, because for some odd reason I couldn't disable HyperThreading even with Apple CHUD installed.
The first machine is an older AMD FX-60 - asus A8N32-SLI board with 2 BFG GTX 280 GPUs (windows 7 64bit).
The second machine is a more up to date MSI brand AM2+ board with an AMD PHENOM II X4 965 black edition and 2 EVGA GTX 275 GPUs (windows 7 64bit).
If I allow machine number 1 with the FX-60 to run ABP2 applications, all I get is a results list full of computation errors. I also suspect that this may be why the machine will blue screen every 2-3 days stating something about a co-processor error.
Machine number 2 will run ABP2 applications all the time without a problem.
Could this be something related to the older AMD socket 939 cpu that I am running, the FX-60?
All other results error out with a "permission denied" sort of error: the app isn't even started in the first place!
Maybe you can check your anti-virus software, sometimes BOINC apps are detected as malware and put it into some sort of quarantine.
There is no known case where a BOINC project distributed malware, but if PC catches a virus, that virus may infect everything on the PC, including BOINC or BOINC apps.
If you are convinced that there is no real problem, you can take the app out of quarantine. By resetting the project in Boinc manager, you can also force BOINC to download fresh copies of the app.
All other results error out with a "permission denied" sort of error: the app isn't even started in the first place!
Maybe you can check your anti-virus software, sometimes BOINC apps are detected as malware and put it into some sort of quarantine.
There is no known case where a BOINC project distributed malware, but if PC catches a virus, that virus may infect everything on the PC, including BOINC or BOINC apps.
If you are convinced that there is no real problem, you can take the app out of quarantine. By resetting the project in Boinc manager, you can also force BOINC to download fresh copies of the app.
CU
HB
Yes, thanks for looking in to this. It was simply that Kaspersky had placed 5-6 different BOINC applications into "untrusted", because if you are not there to answer when Kaspersky asks "is this program OK", it will throw it into the untrusted pile. Took all those out of untrusted and it seems to be ok now. Sorry should have checked this first.
Well i am trying somthing now.
Have a X2 5000+ on it, normaly CPU only.
Installed a 9600GT in it.
Now it does a CPU and a CPU+GPU unit wonder how mutch diff it would take.
Both CPU's are used to 50% and the 9600 stays around 20% usage.
I seem to be stuck running only ABP2 CUDA tasks and haven't had an S5xxx task in months. Does configuring BOINC to run on an NVIDIA GPU exclude running S5xxx tasks?
There are only 10 kind of people in the world, those that understand binary and those that don't!
RE: I've reduced about 75%
)
I'd guess the problem is memory contention. Eight physical cores run E@H tasks can consume a lot of memory bandwidth.
CU
HBE
RE: RE: I've reduced
)
But Nehalem processors have largely addressed the memory contention problem, haven't they...? I used to have a Clovertown-based Mac Pro, that was a real memory contention problem.
For the most part they have.
)
For the most part they have. The 2 channel controller on an quadcore LGA1156 chip can handle 75% of the peak theoretical memory demands of the CPU at DDR3-1333; and that outside of benchmarking it was almost impossible to reach levels high enough to saturate the memory bus. (Intel didn't say exactly which of the chips they were referring to to annanadtech.) The limited bandwidth from the dual channel controller is why they will only be releasing hexacore processors on LGA1366 (Presumably with an uptick in the officially supported ram speed to something above DDR3-1066.)
RE: But Nehalem processors
)
Ah ok, but if we are talking Nehalem CPUs, what about hyperthreading? Did the first, slower, CUDA task runtime correspond to a setup with hyperthreading enabled?
CU
HBE
RE: Ah ok, but if we are
)
This one I wouldn't be sure, because for some odd reason I couldn't disable HyperThreading even with Apple CHUD installed.
I have 2 machines running
)
I have 2 machines running BOINC with dual GPUs.
The first machine is an older AMD FX-60 - asus A8N32-SLI board with 2 BFG GTX 280 GPUs (windows 7 64bit).
The second machine is a more up to date MSI brand AM2+ board with an AMD PHENOM II X4 965 black edition and 2 EVGA GTX 275 GPUs (windows 7 64bit).
If I allow machine number 1 with the FX-60 to run ABP2 applications, all I get is a results list full of computation errors. I also suspect that this may be why the machine will blue screen every 2-3 days stating something about a co-processor error.
Machine number 2 will run ABP2 applications all the time without a problem.
Could this be something related to the older AMD socket 939 cpu that I am running, the FX-60?
RE: Could this be
)
Hi!
I had a look at the errors on this machine. I found one CDUA result that actually worked http://einsteinathome.org/task/160706624.
All other results error out with a "permission denied" sort of error: the app isn't even started in the first place!
Maybe you can check your anti-virus software, sometimes BOINC apps are detected as malware and put it into some sort of quarantine.
There is no known case where a BOINC project distributed malware, but if PC catches a virus, that virus may infect everything on the PC, including BOINC or BOINC apps.
If you are convinced that there is no real problem, you can take the app out of quarantine. By resetting the project in Boinc manager, you can also force BOINC to download fresh copies of the app.
CU
HB
RE: RE: Could this be
)
Yes, thanks for looking in to this. It was simply that Kaspersky had placed 5-6 different BOINC applications into "untrusted", because if you are not there to answer when Kaspersky asks "is this program OK", it will throw it into the untrusted pile. Took all those out of untrusted and it seems to be ok now. Sorry should have checked this first.
Well i am trying somthing
)
Well i am trying somthing now.
Have a X2 5000+ on it, normaly CPU only.
Installed a 9600GT in it.
Now it does a CPU and a CPU+GPU unit wonder how mutch diff it would take.
Both CPU's are used to 50% and the 9600 stays around 20% usage.
XP x64
2GB DDR 400
Nv 650i
I seem to be stuck running
)
I seem to be stuck running only ABP2 CUDA tasks and haven't had an S5xxx task in months. Does configuring BOINC to run on an NVIDIA GPU exclude running S5xxx tasks?
There are only 10 kind of people in the world, those that understand binary and those that don't!