There are special app versions that we are testing right now, specifically compiled and tuned for the CPUs we do have many of in the ATLAS computing cluster. These are plan classes "ATLAS1" and " "ATLAS2", and are bound to CPU models "Intel(R) Xeon(R) CPU E3-1220 v3" and "Intel(R) Xeon(R) CPU E5-2658 v4" respectively. These are not restricted to ATLAS nodes, though, if you happen to have the same CPU, you'll get these too.
(this can take several minutes or even hours) on your machine and put the resulting file FGRPB1wisdom.dat in the E@H project directory (I think the standard path is /var/lib/boinc/projects/einstein.phys.uwm.edu). "rib" should be enough. You may even get an improvement with "measure" "wisdom" (add "--measure" to the command-line); generation will be faster but improvement not so large.
Apparently the switch from FFTW-3.3.2 to 3.3.6 wasn't beneficial for all machines. It looks like the new app version is notably slower than the previous 1.5 on some machines, on average it's almost on par.
Hm, the stderr is cut off at the top (which I think is common). Do you happen to know whether there's a message saying "Successfully read wisdom from file" in the log shortly after the start (precisely: after the "Creating FFT ... plan" message, as in that task log?
With FGRPSSE version 1.05 I never had invalids (as far as I remember), but recently, since versions 1.06 and 1.07 are out, I noticed several 1.05 invalids - just to mention that fact.
There are special app
)
There are special app versions that we are testing right now, specifically compiled and tuned for the CPUs we do have many of in the ATLAS computing cluster. These are plan classes "ATLAS1" and " "ATLAS2", and are bound to CPU models "Intel(R) Xeon(R) CPU E3-1220 v3" and "Intel(R) Xeon(R) CPU E5-2658 v4" respectively. These are not restricted to ATLAS nodes, though, if you happen to have the same CPU, you'll get these too.
BM
Here's a link to
)
Here's fftwf-3.3.6-pl2-Linux64-wisdom, the fftwf-wisdom binary that should produce wisdom that can be imported by the 1.06 application. Run
(this can take several minutes or even hours) on your machine and put the resulting file FGRPB1wisdom.dat in the E@H project directory (I think the standard path is /var/lib/boinc/projects/einstein.phys.uwm.edu). "rib" should be enough. You may even get an improvement with "measure" "wisdom" (add "--measure" to the command-line); generation will be faster but improvement not so large.
BM
Apparently the switch from
)
Apparently the switch from FFTW-3.3.2 to 3.3.6 wasn't beneficial for all machines. It looks like the new app version is notably slower than the previous 1.5 on some machines, on average it's almost on par.
BM
Bernd Machenschalk
)
It should be: rib67108864
Yes, it should. Sorry, fixed
)
Yes, it should. Sorry, fixed by editing.
BM
I run that wisdom-thing and
)
I run that wisdom-thing and first five tasks that read it have completed now. They took the same time as earlier tasks without that.
Hm, the stderr is cut off at
)
Hm, the stderr is cut off at the top (which I think is common). Do you happen to know whether there's a message saying "Successfully read wisdom from file" in the log shortly after the start (precisely: after the "Creating FFT ... plan" message, as in that task log?
BM
Me? Yes, I made that change
)
Me? Yes, I made that change for my host 12542048 so that all tasks that have uploaded after the beginning of 28 Jun include this in the Stderr:
With FGRPSSE version 1.05 I
)
With FGRPSSE version 1.05 I never had invalids (as far as I remember), but recently, since versions 1.06 and 1.07 are out, I noticed several 1.05 invalids - just to mention that fact.
I have two windows 1.05
)
I have two windows 1.05 workunits failing to validate against 1.06/1.07 workunits:
https://einsteinathome.org/workunit/303346295
https://einsteinathome.org/workunit/303097468