Good results with idle=poll + help needed
22 Sep 2014 03:41 #51411
by peter76
Good results with idle=poll + help needed was created by peter76
Hello,
I'm fiddling with a MSI K9N6PGM2-V board. Booting with isolcpus=1 gives me about 9500 us jitter in the base thread.
If I also boot with idle=poll, the jitter drops to around 4200 (!).
Problem is this is wasting energy and googling around learned it is possible to do this on demand instead of on boot, by writing the desired latency to /dev/cpu_dma_latency.
See here: access.redhat.com/articles/65410 and here: software.intel.com/en-us/forums/topic/392866.
I tried the given examples, but they don't work and my C knowledge is just a but more than Hello World!
Anybody here that can make one of the programs to work, so that it is possible to only have the cpu's constant attention when running LinuxCnc.
( Sorry if I'm a bit vague, reading the Red Hat link will make things clearer )
I'm also interested what results others are getting booting with idle=poll.
Regards, Peter
I'm fiddling with a MSI K9N6PGM2-V board. Booting with isolcpus=1 gives me about 9500 us jitter in the base thread.
If I also boot with idle=poll, the jitter drops to around 4200 (!).
Problem is this is wasting energy and googling around learned it is possible to do this on demand instead of on boot, by writing the desired latency to /dev/cpu_dma_latency.
See here: access.redhat.com/articles/65410 and here: software.intel.com/en-us/forums/topic/392866.
I tried the given examples, but they don't work and my C knowledge is just a but more than Hello World!
Anybody here that can make one of the programs to work, so that it is possible to only have the cpu's constant attention when running LinuxCnc.
( Sorry if I'm a bit vague, reading the Red Hat link will make things clearer )
I'm also interested what results others are getting booting with idle=poll.
Regards, Peter
Please Log in or Create an account to join the conversation.
22 Sep 2014 04:38 #51414
by peter76
Replied by peter76 on topic Good results with idle=poll + help needed
PS
Now also applied the irq stuff from this thread:
www.linuxcnc.org/index.php/english/forum...ulticore-pcs-success
Having 3 glxgears runnin and getting:
1606
I'm havng fun:-)
Regards, Peter
Now also applied the irq stuff from this thread:
www.linuxcnc.org/index.php/english/forum...ulticore-pcs-success
Having 3 glxgears runnin and getting:
1606
I'm havng fun:-)
Regards, Peter
Please Log in or Create an account to join the conversation.
23 Sep 2014 14:51 #51477
by ArcEye
We tested this and various other kernel switches during the development of ubc branch of Linuxcnc, which later became MachineKit.
It certainly had good effect with xenomai kernels.
I think you will find there is a law of diminishing returns comes into effect.
It is possible to get the latency figure down very low, by crippling the processor and allowing only realtime tasks, but the other side of this is that userspace programs, which include Axis etc, run like their legs have been cut off.
A latency base thread figure of 5 - 9K is very good for software stepping and should allow the rest of the software to work properly too.
regards
Replied by ArcEye on topic Good results with idle=poll + help needed
I'm also interested what results others are getting booting with idle=poll.
We tested this and various other kernel switches during the development of ubc branch of Linuxcnc, which later became MachineKit.
It certainly had good effect with xenomai kernels.
I think you will find there is a law of diminishing returns comes into effect.
It is possible to get the latency figure down very low, by crippling the processor and allowing only realtime tasks, but the other side of this is that userspace programs, which include Axis etc, run like their legs have been cut off.
A latency base thread figure of 5 - 9K is very good for software stepping and should allow the rest of the software to work properly too.
regards
Please Log in or Create an account to join the conversation.
30 Sep 2014 02:48 #51671
by peter76
Replied by peter76 on topic Good results with idle=poll + help needed
Ok, I just ran the latency test and was quite happy, haven't tried to run Axis yet with the idle=poll option....
Still would like to know why writing to /dev/cpu_dma_latency doesn't work.
regards, Peter
Still would like to know why writing to /dev/cpu_dma_latency doesn't work.
regards, Peter
Please Log in or Create an account to join the conversation.
Time to create page: 0.104 seconds