I've watched my host that showed the affinity loss issue more closely. Here is what I observe.
Following boot, the work is split evenly (50% or less each) between the two (logical) cores. (The machine is a Pentium 4 with HT). After some period of time, apparently days, something "gives way" and the affinity is lost AND STAYS LOST. If I reboot the machine, the affinity is restored when the two E@H jobs resume. This has been observed two or three times since my last post.
Other apps that run on the machine: some limited web browsing, "text file to MP3 speech file" processing (Text-Aloud), and sidebar apps (weather, time, date and computer load).
I've watched my host that showed the affinity loss issue more closely. Here is what I observe.
Following boot, the work is split evenly (50% or less each) between the two (logical) cores. (The machine is a Pentium 4 with HT). After some period of time, apparently days, something "gives way" and the affinity is lost AND STAYS LOST. If I reboot the machine, the affinity is restored when the two E@H jobs resume. This has been observed two or three times since my last post.
Other apps that run on the machine: some limited web browsing, "text file to MP3 speech file" processing (Text-Aloud), and sidebar apps (weather, time, date and computer load).
The machine identification was incorrect ... the machine is a Pentium D 3.0 GHz running Vista Home Premium SP1 (32-bit version).
Sorry for the misinfo ... my brain was "on vacation."
But is CPU affinity really an issue at all with single CPU hyperthreading systems? As I understand it, the real purpose behind CPU affinity is "cache affinity". But with hyperthreading, all cache levels are shared for all logical CPUs of a hyperthreading CPU.
I've read much of this thread in hopes of finding a quick and easy answer. Hmmm. To no avail indeed: much about modifying boinc, etc. A bit much for someone who knows little on such matters. I seem to have the same problem: Einstein just doesn't run right with two or more WUs running simultaneously. The clock on the CPU runs at half speed, and the core usage is not at all at 100% as is usually the case. I note that the memory usage is still where it should be when running other projects. So, what do I do?
I've read much of this thread in hopes of finding a quick and easy answer. Hmmm. To no avail indeed: much about modifying boinc, etc. A bit much for someone who knows little on such matters. I seem to have the same problem: Einstein just doesn't run right with two or more WUs running simultaneously. The clock on the CPU runs at half speed, and the core usage is not at all at 100% as is usually the case. I note that the memory usage is still where it should be when running other projects. So, what do I do?
If your cpu is automatically dropping back its clock speed when running BOINC I would suspect it is overheating. I just compared your times with my host 252515 and the news is not good.
I have no problems running any combination of Einstein, incl power apps, and Seti using optimised apps, Temperatures never higher than 60C, even on the warmest days.
If your cpu is automatically dropping back its clock speed when running BOINC I would suspect it is overheating. I just compared your times with my host 252515 and the news is not good.
I have no problems running any combination of Einstein, incl power apps, and Seti using optimised apps, Temperatures never higher than 60C, even on the warmest days.
Heat doesn't appear to be the real problem. I have three hosts, all of which run BOINC 24/7 with no heat issues. Been there, done that. The problem exists exclusively when I run Einstein WUs. When Einstein runs, CPU usage drops off dramatically, yet all other projects run without problems, in about the time expected. I do note that my CPU times are two to three times what your identical host does with these current WUs. This is not positive!
But is CPU affinity really an issue at all with single CPU hyperthreading systems? As I understand it, the real purpose behind CPU affinity is "cache affinity". But with hyperthreading, all cache levels are shared for all logical CPUs of a hyperthreading CPU.
CU
Bikeman
I'd have to agree here. Affinity is an interesting phenomena to observe and play around with.
However, I always thought that it was only really worth fooling with for a dedicated crunching rig. My feeling is your point applies to physical cores as well as logical ones.
On gear with a 'day job', and as software becomes more demanding and multi-core aware going forward, it seems to me that a BOINC app is more likely to get kicked out of cache (at least the ones closest to the execution units) even if it is 'locked' to a particular core. Granted, you would still save the overhead involved with a full context switch if the task got moved to a different core, but how often is that going to happen if you are allowing all cores to execute BOINC in the first place?
From what I read, that could lead to inefficiencies being introduced for the hosts primary function when the OS'es task scheduler has to make allowances for a process that wants to cling to to a core rather than be able to allocate resources with the overall picture in mind. The papers presented about how hyperthreading actually was actually slowing down performance in some database applications with the first generation implementation comes to mind here.
It will be interesting to see how that changes with HT2 in Nehalem.
I've read much of this thread in hopes of finding a quick and easy answer. Hmmm. To no avail indeed: much about modifying boinc, etc. A bit much for someone who knows little on such matters. I seem to have the same problem: Einstein just doesn't run right with two or more WUs running simultaneously. The clock on the CPU runs at half speed, and the core usage is not at all at 100% as is usually the case. I note that the memory usage is still where it should be when running other projects. So, what do I do?
I've had a few WUs download since my first post above to this thread, and the host continued to have some problems. The one thing I didn't try to get my processor processing at 100% was do a detach/reattach from E@h. I even tried a project reset to no avail. Sure enough, tried it and thar she blows: E@h now runs just the way it should. Still not sure what it did, but things must have needed to get cleaned out and started from scratch. Hope this helps others with affinity problems.
I have the following configuration and do not have this problem.
Mother board Asus P5K-E WiFi AP
CPU type GenuineIntel
Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz [Intel64 Family 6 Model 15 Stepping 11]
Number of CPUs 4
Operating System Microsoft Windows Vista
Home Premium x64 Editon, Service Pack 1, (06.00.6001.00)
Memory 4094.18 MB
Cache 244.14 KB
Swap space 8379.65 MB
Total disk space 10.74 GB
Free Disk Space 9.66 GB
I have, after reading a lot of posts regarding over clocking and stability, done the following.
In BIOS under the CPU menu I have manually disable the following.
CIE support
Max CPUID Value limit
Vanderpool Technology
CPU TM Function
Execute Disable Bit
I've watched my host that
)
I've watched my host that showed the affinity loss issue more closely. Here is what I observe.
Following boot, the work is split evenly (50% or less each) between the two (logical) cores. (The machine is a Pentium 4 with HT). After some period of time, apparently days, something "gives way" and the affinity is lost AND STAYS LOST. If I reboot the machine, the affinity is restored when the two E@H jobs resume. This has been observed two or three times since my last post.
Other apps that run on the machine: some limited web browsing, "text file to MP3 speech file" processing (Text-Aloud), and sidebar apps (weather, time, date and computer load).
Stan
I said: RE: I've watched
)
I said:
The machine identification was incorrect ... the machine is a Pentium D 3.0 GHz running Vista Home Premium SP1 (32-bit version).
Sorry for the misinfo ... my brain was "on vacation."
Stan
But is CPU affinity really an
)
But is CPU affinity really an issue at all with single CPU hyperthreading systems? As I understand it, the real purpose behind CPU affinity is "cache affinity". But with hyperthreading, all cache levels are shared for all logical CPUs of a hyperthreading CPU.
CU
Bikeman
I've read much of this thread
)
I've read much of this thread in hopes of finding a quick and easy answer. Hmmm. To no avail indeed: much about modifying boinc, etc. A bit much for someone who knows little on such matters. I seem to have the same problem: Einstein just doesn't run right with two or more WUs running simultaneously. The clock on the CPU runs at half speed, and the core usage is not at all at 100% as is usually the case. I note that the memory usage is still where it should be when running other projects. So, what do I do?
Windows Vista SP1
Q6600 Core 2 Quad
Here's the Host.
(Click for detailed stats)
RE: I've read much of this
)
If your cpu is automatically dropping back its clock speed when running BOINC I would suspect it is overheating. I just compared your times with my host 252515 and the news is not good.
I have no problems running any combination of Einstein, incl power apps, and Seti using optimised apps, Temperatures never higher than 60C, even on the warmest days.
RE: If your cpu is
)
Heat doesn't appear to be the real problem. I have three hosts, all of which run BOINC 24/7 with no heat issues. Been there, done that. The problem exists exclusively when I run Einstein WUs. When Einstein runs, CPU usage drops off dramatically, yet all other projects run without problems, in about the time expected. I do note that my CPU times are two to three times what your identical host does with these current WUs. This is not positive!
(Click for detailed stats)
RE: But is CPU affinity
)
I'd have to agree here. Affinity is an interesting phenomena to observe and play around with.
However, I always thought that it was only really worth fooling with for a dedicated crunching rig. My feeling is your point applies to physical cores as well as logical ones.
On gear with a 'day job', and as software becomes more demanding and multi-core aware going forward, it seems to me that a BOINC app is more likely to get kicked out of cache (at least the ones closest to the execution units) even if it is 'locked' to a particular core. Granted, you would still save the overhead involved with a full context switch if the task got moved to a different core, but how often is that going to happen if you are allowing all cores to execute BOINC in the first place?
From what I read, that could lead to inefficiencies being introduced for the hosts primary function when the OS'es task scheduler has to make allowances for a process that wants to cling to to a core rather than be able to allocate resources with the overall picture in mind. The papers presented about how hyperthreading actually was actually slowing down performance in some database applications with the first generation implementation comes to mind here.
It will be interesting to see how that changes with HT2 in Nehalem.
Alinator
RE: I've read much of this
)
I've had a few WUs download since my first post above to this thread, and the host continued to have some problems. The one thing I didn't try to get my processor processing at 100% was do a detach/reattach from E@h. I even tried a project reset to no avail. Sure enough, tried it and thar she blows: E@h now runs just the way it should. Still not sure what it did, but things must have needed to get cleaned out and started from scratch. Hope this helps others with affinity problems.
(Click for detailed stats)
I have the following
)
I have the following configuration and do not have this problem.
Mother board Asus P5K-E WiFi AP
CPU type GenuineIntel
Intel(R) Core(TM)2 Quad CPU Q6600 @ 2.40GHz [Intel64 Family 6 Model 15 Stepping 11]
Number of CPUs 4
Operating System Microsoft Windows Vista
Home Premium x64 Editon, Service Pack 1, (06.00.6001.00)
Memory 4094.18 MB
Cache 244.14 KB
Swap space 8379.65 MB
Total disk space 10.74 GB
Free Disk Space 9.66 GB
I have, after reading a lot of posts regarding over clocking and stability, done the following.
In BIOS under the CPU menu I have manually disable the following.
CIE support
Max CPUID Value limit
Vanderpool Technology
CPU TM Function
Execute Disable Bit
Hope this helps.