Hmm, I don't think that deciding based on CPU name is a good idea.
What about:
Xeon E5-2xxx
Xeon E5-4xxx
Xeon E7-xxxx
Xeon D-xxxx
Intel(R) Core(TM)2 Quad CPU Q9650
Core i3-6320 has 4M cache
etc...
Also, there are sample CPUs, which don't contain a branding string and still can be very powerful: "Genuine Intel(R) CPU @ 3.10GHz" - a Xeon sample.
Yeah, incredible that all those high end retail xeon CPUs are left out of processing these units.
Other powerful xeon E5's engineering samples CPU string names are:
Genuine Intel(R) CPU E2697V @ 2.70GHz [Family 6 Model 62 Stepping 4]
GenuineIntel 06/2d [Family 6 Model 45 Stepping 6]
These ones are in two of my hosts :), C1 stepping so preproduction units.
Yeah, I'm wondering what the exact "test" is, because my piddly little HP Microserver G8 with a Xeon E3-1220L V2 (and it's corresponding piddly 3MB L3 cache) just got 4 of the "F" tasks.
If a 3MB L3 is good enough, then just who exactly is getting the "I"s? ;-)
Edit: Nevermind... I just read down and apparently 3MB is the "cutoff", so I guess we'll just see what a whopping 17W TDP can do!
I'd suggest a better method based on CPUID output, but since the decision making code is run on the server side, it needs to rely on BOINC output and thus probably needs to use the CPU name string.
Or let the user decide...
- We are currently starting the "actual" run that was announced a few minutes ago. It will show up as two different applications that will be sent to different groups of hosts (O1AS20-100F, think of "F" as "fast hosts", and O1AS20-100I).
Thanks for the update Bernd, and the new apps. Iḿ picking up F apps on one host perfect.
I´m thinking this i3 530 host should pick up ¨X64O1I" but scheduler says ... no to F tasks
2016-03-09 21:04:25.4611 [PID=31986] [version] Checking plan class 'AVXO1F'
2016-03-09 21:04:25.4635 [PID=31986] [version] reading plan classes from file '/BOINC/projects/EinsteinAtHome/plan_class_spec.xml'
2016-03-09 21:04:25.4636 [PID=31986] [version] CPU [' fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx rdtscp lm constant_tsc arch_perfmon pebs bts rep_good xtopology nonstop_tsc aperfmperf pni dtes64 monitor ds_cpl vm family 6 model 37 stepping 2 '] lacks feature ' avx '
2016-03-09 21:04:25.4636 [PID=31986] [version] Checking plan class 'X64O1F'
2016-03-09 21:04:25.4636 [PID=31986] [version] CPU model 'Intel(R) Core(TM) i3 CPU 530 @ 2.93GHz [Family 6 Model 37 Stepping 2]' doesn't match 'I3-3225| i[57][- ]| E[38]| E5-1| X[57]| G[238]| G64'
2016-03-09 21:04:25.4637 [PID=31986] [version] Checking plan class 'SSE2O1F'
2016-03-09 21:04:25.4637 [PID=31986] [version] CPU model 'Intel(R) Core(TM) i3 CPU 530 @ 2.93GHz [Family 6 Model 37 Stepping 2]' doesn't match 'I3-3225| i[57][- ]| E[38]| E5-1| X[57]| G[238]| G64'
2016-03-09 21:04:25.4637 [PID=31986] [version] no app version available: APP#35 (einstein_O1AS20-100F) PLATFORM#7 (x86_64-pc-linux-gnu) min_version 0
2016-03-09 21:04:25.4637 [PID=31986] [version] no app version available: APP#35 (einstein_O1AS20-100F) PLATFORM#1 (i686-pc-linux-gnu) min_version 0
This is the first time we have two "locality scheduling" applications on Einstein@Home and there are still a few oddities. Please bear with us while we are ironing out the remaining glitches.
One issue is that the ("locality") scheduler first picks a task to send to a host, then tries to find an application version to process it. If there is no suitable app version for the host the scheduler just plainly gives up instead of picking another task.
I just tweaked the scheduler code a bit, narrowing down the selection of tasks for a given host. This should fix most of the problems.
Phenom II series CPUs have 6 MB L3 cache and FX-8xxx/FX-6xxx series have 8 MB L3
But seems getting only I-series WUs
Also AVX detection seems not working properly - my FX-8320 hosts getting X64O1I app now, while in "tuning" run they run AVX app. Or AVX and SSE2-64bit apps was merged and now detect CPU specs internally?
BTW - points per task has been doubled? 2000 Cr per "O1 all-sky I" task vs 1000 Cr per "O1 all-sky tuning" task. While compute time is about at same scale (may be increased but not much).
RE: Hmm, I don't think
)
Yeah, incredible that all those high end retail xeon CPUs are left out of processing these units.
Other powerful xeon E5's engineering samples CPU string names are:
Genuine Intel(R) CPU E2697V @ 2.70GHz [Family 6 Model 62 Stepping 4]
GenuineIntel 06/2d [Family 6 Model 45 Stepping 6]
These ones are in two of my hosts :), C1 stepping so preproduction units.
RE: Hmm, I don't think that
)
the project has just started. I think that this list will be expanded.
How many people can have e5,e7??
i5,i7 - for volunteers. E5,e7 and etc - for clusters. ;)
ps. Quad CPU- is very good, but 24+ hour for each WU is not good.
ps2. HT is very bad for O1AS20-100F. i7-3770 (8 Threads) <= i5-3470
In our team at XtremeSystems,
)
In our team at XtremeSystems, E5's are not uncommon ;-) http://www.xtremesystems.org/forums/showthread.php?294437-DC-PrOn&p=5272650&viewfull=1#post5272650
-----
Yeah, I'm wondering what the
)
Yeah, I'm wondering what the exact "test" is, because my piddly little HP Microserver G8 with a Xeon E3-1220L V2 (and it's corresponding piddly 3MB L3 cache) just got 4 of the "F" tasks.
If a 3MB L3 is good enough, then just who exactly is getting the "I"s? ;-)
Edit: Nevermind... I just read down and apparently 3MB is the "cutoff", so I guess we'll just see what a whopping 17W TDP can do!
I'd suggest a better method
)
I'd suggest a better method based on CPUID output, but since the decision making code is run on the server side, it needs to rely on BOINC output and thus probably needs to use the CPU name string.
Or let the user decide...
-----
RE: - We are currently
)
Thanks for the update Bernd, and the new apps. Iḿ picking up F apps on one host perfect.
I´m thinking this i3 530 host should pick up ¨X64O1I" but scheduler says ... no to F tasks
but does not look for I.
I have tried manually setting the preferences
But i still don´t see ¨Checking plan class¨ for the I app.
hth.
This i7 Win8.1 host is
)
This i7 Win8.1 host is picking up 'F' tasks, but currently not this i7 Fedora host.
Cheers, Mike.
( edit ) The Fedora has just got a bunch of 'F' units. Cool groover ... this is exciting ! :-)
I have made this letter longer than usual because I lack the time to make it shorter ...
... and my other CPU is a Ryzen 5950X :-) Blaise Pascal
This is the first time we
)
This is the first time we have two "locality scheduling" applications on Einstein@Home and there are still a few oddities. Please bear with us while we are ironing out the remaining glitches.
One issue is that the ("locality") scheduler first picks a task to send to a host, then tries to find an application version to process it. If there is no suitable app version for the host the scheduler just plainly gives up instead of picking another task.
I just tweaked the scheduler code a bit, narrowing down the selection of tasks for a given host. This should fix most of the problems.
BM
BM
You forgot about AMD
)
You forgot about AMD CPUs?
Phenom II series CPUs have 6 MB L3 cache and FX-8xxx/FX-6xxx series have 8 MB L3
But seems getting only I-series WUs
Also AVX detection seems not working properly - my FX-8320 hosts getting X64O1I app now, while in "tuning" run they run AVX app. Or AVX and SSE2-64bit apps was merged and now detect CPU specs internally?
BTW - points per task has been doubled? 2000 Cr per "O1 all-sky I" task vs 1000 Cr per "O1 all-sky tuning" task. While compute time is about at same scale (may be increased but not much).
is the 64 bit AVX also ?
)
is the 64 bit AVX also ?