I'm not sure if you restarted boinc or not after recalculating, but in order to get a decision you need to let it settle into a steady state. I think the wisdom is only read once when the task starts (or is restarted). So changing windom when running will lead to new tasks running with a different wisdom. You can see that in the graph where the second wisdom tasks overtake the first wisdom tasks.
Even if it runs slower - you can use that loading to create a second wisdom and try that. That lead, for me, to ~10% faster results.
I'm not sure if you restarted boinc or not after recalculating, but in order to get a decision you need to let it settle into a steady state. I think the wisdom is only read once when the task starts (or is restarted).
I didn't restart it after "wisdom 2" had been generated. I wanted this thing to change on-the-fly, so that the old tasks completed with wisdom 1 and then the new tasks start with wisdom 2.
Quote:
So changing windom when running will lead to new tasks running with a different wisdom.
There are several bottlenecks, how fast memory can be moved in and out of cache(s), processor speed, size of SIMD registrars and probably a bunch more.
I only really tested on the Xeon, and if i had some more time i'd start up PCM and see what is really going on (with each wisdom).
.
After crunching some more of those 1.08s on a Skylake, some with wisdom generated at all engines full stop condition, some with wisdom generated at 90% CPU load condition, I didn't notice any obvious difference. Thus that whole procedure to me seems to be pretty robust.
Another useful tool might be Intel VTune, if you download the 30 day trial version. Instead, I ran that little perfmon tool, which gathers several data for 60 seconds, then reports. I noticed that while workunits were running that had read the heavy load wisdom, the CPU section of the report said that the swap file was some 10 percent smaller than in the other case. However, I don't know whether that's significant, neither whether that's a result of a wisdom instruction, and I doubt it is relevant at all for this machine, as it didn't show up in the overall CPU time. So I guess it won't be easy to determine a bottleneck with a tool, and we'll have to rely upon the wisdom to do the best job.
I noticed that while workunits were running that had read the heavy load wisdom, the CPU section of the report said that the swap file was some 10 percent smaller than in the other case.
I noticed quite a difference in memory usage, one inefficient wisdom required 53% more memory (510 MB to 772 MB peak size).
I just read about the wisdom-files and the possibility to speed up CPU-Crunching. So I tried to create a wisdom file on my cruncher that's running on Linux (Mint). Simply opened the project-folder where there is a file called "fftwf-wisdom_FGRP5_1.08_x86_64-pc-linux-gnu". Opened the console and typed the following
I'm not sure if you restarted
)
I'm not sure if you restarted boinc or not after recalculating, but in order to get a decision you need to let it settle into a steady state. I think the wisdom is only read once when the task starts (or is restarted). So changing windom when running will lead to new tasks running with a different wisdom. You can see that in the graph where the second wisdom tasks overtake the first wisdom tasks.
Even if it runs slower - you can use that loading to create a second wisdom and try that. That lead, for me, to ~10% faster results.
Good luck in search of wisdom...
AgentB wrote:I'm not sure if
)
I didn't restart it after "wisdom 2" had been generated. I wanted this thing to change on-the-fly, so that the old tasks completed with wisdom 1 and then the new tasks start with wisdom 2.
Yes, that was what I aimed for.
Thanks!
AgentB schrieb: There are
)
After crunching some more of those 1.08s on a Skylake, some with wisdom generated at all engines full stop condition, some with wisdom generated at 90% CPU load condition, I didn't notice any obvious difference. Thus that whole procedure to me seems to be pretty robust.
Another useful tool might be Intel VTune, if you download the 30 day trial version. Instead, I ran that little perfmon tool, which gathers several data for 60 seconds, then reports. I noticed that while workunits were running that had read the heavy load wisdom, the CPU section of the report said that the swap file was some 10 percent smaller than in the other case. However, I don't know whether that's significant, neither whether that's a result of a wisdom instruction, and I doubt it is relevant at all for this machine, as it didn't show up in the overall CPU time. So I guess it won't be easy to determine a bottleneck with a tool, and we'll have to rely upon the wisdom to do the best job.
solling2 wrote: I noticed
)
I noticed quite a difference in memory usage, one inefficient wisdom required 53% more memory (510 MB to 772 MB peak size).
Hey guys, I just read
)
Hey guys,
I just read about the wisdom-files and the possibility to speed up CPU-Crunching. So I tried to create a wisdom file on my cruncher that's running on Linux (Mint). Simply opened the project-folder where there is a file called "fftwf-wisdom_FGRP5_1.08_x86_64-pc-linux-gnu". Opened the console and typed the following
fftwf-wisdom_FGRP5_1.08_x86_64-pc-linux-gnu -o FGRPB1wisdom.dat rib67108864
It just says, the instruction can't be executed. What am I doing wrong?!?
It's a noob-question - sorry for that.
Is the file marked
)
Is the file marked executable? What you did in Terminal is all that is needed to generate the FGRPB1wisdom.dat file.
In the GUI the file is shown
)
In the GUI the file is shown as "executable".
Just to proof it, I tried:
sudo chmod +x fftwf-wisdom_FGRP5_1.08_x86_64-pc-linux-gnu
The Command seems to work; after hitting 'enter' the console doesn't show any messages, so I think the instruction is done without any failures.
Trying
fftwf-wisdom_FGRP5_1.08_x86_64-pc-linux-gnu -o FGRPB1wisdom.dat rib67108864
it says again the instruction can't be executed.
Ohh, should have paid more
)
Ohh, should have paid more attention to the filename you are using. You are not running the correct file. It should be:
fftwf-wisdom_FGRPB1_1.08_x86_64-pc-linux-gnu -o FGRPB1wisdom.dat rib67108864
[Edit]
Ignore my last. You are asking about the cpu wisdom file. Correct?
Maybe it some basics? Do you
)
Maybe it some basics? Do you prepended command with ./ ?
./fftwf-wisdom_FGRP5_1.08_x86_64-pc-linux-gnu -o FGRPB1wisdom.dat rib67108864
Keith Myers schrieb: Ohh,
)
Yeah I'm just asking for the CPU wisdom file. Copied the filename in the GUI and pasted it in the console, so I think the name is right.
Are there wisdom-files for GPU, too?