I have a multi processor system/machine and I would like dedicate a CPU entirely to listen on to a port. I would greatly appreciate if you could tell me how this can be done.
Thanks!
On Fri, Sep 16, 2011 at 5:08 AM, Kernel Piddu <kernelpiddu@gmail.com> wrote:
I have a multi processor system/machine and I would like dedicate a CPU entirely to listen on to a port. I would greatly appreciate if you could tell me how this can be done.
Thanks!
doing this in the kernel, is perhaps not really recommended, and linux kernel is become UNIX-non-compliant if such features were ever provided. but userspace, is definitely possible, but is always a suggestive tuning....the kernel can always override your request: http://www.cyberciti.biz/tips/setting-processor-affinity-certain-task-or-pro... but this is per-process, if u want per-port IRQ handling attach to the core, then this is going to affect the algorithm kernel used for SoftIRQ handling: doing a "cat /proc/softirqs" on my system: CPU0 CPU1 CPU2 CPU3 HI: 0 0 0 0 TIMER: 563854 591025 0 0 NET_TX: 7 6 0 0 NET_RX: 17486 19250 0 0 BLOCK: 28888 3784 0 0 BLOCK_IOPOLL: 0 0 0 0 TASKLET: 399230 391008 0 0 SCHED: 116858 119604 0 0 HRTIMER: 978 610 0 0 RCU: 158513 192567 0 0 i can see that the IRQ processing is skewed, but then unless i have an overall picture of what every core is doing.....not advisable to touch that. anyway looking at the "processes": root 3 2 0 16:59 ? 00:00:00 [ksoftirqd/0] root 7 2 0 16:59 ? 00:00:00 [ksoftirqd/1] only two tasklet thread are spawned to handle the these softirq at each core. i suspect some error in the interpreting the number of cores here, as there is only two core. (/proc/cpuinfo). good reading on these internals: http://lwn.net/Articles/308117/ http://www.wil.cx/matthew/lca2003/paper.pdf -- Regards, Peter Teoh
Thanks so much for the detailed info. It's very comprehensive. Thanks for taking the time. On Sep 26, 2011, at 6:51 AM, Peter Teoh <htmldeveloper@gmail.com> wrote:
On Fri, Sep 16, 2011 at 5:08 AM, Kernel Piddu <kernelpiddu@gmail.com> wrote:
I have a multi processor system/machine and I would like dedicate a CPU entirely to listen on to a port. I would greatly appreciate if you could tell me how this can be done.
Thanks!
doing this in the kernel, is perhaps not really recommended, and linux kernel is become UNIX-non-compliant if such features were ever provided. but userspace, is definitely possible, but is always a suggestive tuning....the kernel can always override your request: http://www.cyberciti.biz/tips/setting-processor-affinity-certain-task-or-pro... but this is per-process, if u want per-port IRQ handling attach to the core, then this is going to affect the algorithm kernel used for SoftIRQ handling:
doing a "cat /proc/softirqs" on my system:
CPU0 CPU1 CPU2 CPU3 HI: 0 0 0 0 TIMER: 563854 591025 0 0 NET_TX: 7 6 0 0 NET_RX: 17486 19250 0 0 BLOCK: 28888 3784 0 0 BLOCK_IOPOLL: 0 0 0 0 TASKLET: 399230 391008 0 0 SCHED: 116858 119604 0 0 HRTIMER: 978 610 0 0 RCU: 158513 192567 0 0
i can see that the IRQ processing is skewed, but then unless i have an overall picture of what every core is doing.....not advisable to touch that.
anyway looking at the "processes":
root 3 2 0 16:59 ? 00:00:00 [ksoftirqd/0] root 7 2 0 16:59 ? 00:00:00 [ksoftirqd/1]
only two tasklet thread are spawned to handle the these softirq at each core. i suspect some error in the interpreting the number of cores here, as there is only two core. (/proc/cpuinfo).
good reading on these internals:
http://lwn.net/Articles/308117/ http://www.wil.cx/matthew/lca2003/paper.pdf
-- Regards, Peter Teoh
well, use the term "cpu affinity" when u google and u will find more info (correct) from others - my answer is not complete - there are situation where u can do tuning - read further here: http://honglus.blogspot.com/2011/03/tune-interrupt-and-process-cpu-affinity.... http://www.ibm.com/developerworks/linux/library/l-affinity/index.html at the lowest level, the IRQ is handled at the the CPU inside arch/x86/kernel/irq_32.c. Inside this file, u can always identify your own SMP core by a specific smp ID, and if it matches, then continue processing the IRQ, otherwise, just ignore the processing, and let the other processor does the work - not sure if I am right? On Mon, Sep 26, 2011 at 7:28 PM, Kernel Piddu <kernelpiddu@gmail.com> wrote:
Thanks so much for the detailed info. It's very comprehensive. Thanks for taking the time.
On Sep 26, 2011, at 6:51 AM, Peter Teoh <htmldeveloper@gmail.com> wrote:
On Fri, Sep 16, 2011 at 5:08 AM, Kernel Piddu <kernelpiddu@gmail.com> wrote:
I have a multi processor system/machine and I would like dedicate a CPU entirely to listen on to a port. I would greatly appreciate if you could tell me how this can be done.
Thanks!
doing this in the kernel, is perhaps not really recommended, and linux kernel is become UNIX-non-compliant if such features were ever provided. but userspace, is definitely possible, but is always a suggestive tuning....the kernel can always override your request: http://www.cyberciti.biz/tips/setting-processor-affinity-certain-task-or-pro... but this is per-process, if u want per-port IRQ handling attach to the core, then this is going to affect the algorithm kernel used for SoftIRQ handling:
doing a "cat /proc/softirqs" on my system:
CPU0 CPU1 CPU2 CPU3 HI: 0 0 0 0 TIMER: 563854 591025 0 0 NET_TX: 7 6 0 0 NET_RX: 17486 19250 0 0 BLOCK: 28888 3784 0 0 BLOCK_IOPOLL: 0 0 0 0 TASKLET: 399230 391008 0 0 SCHED: 116858 119604 0 0 HRTIMER: 978 610 0 0 RCU: 158513 192567 0 0
i can see that the IRQ processing is skewed, but then unless i have an overall picture of what every core is doing.....not advisable to touch that.
anyway looking at the "processes":
root 3 2 0 16:59 ? 00:00:00 [ksoftirqd/0] root 7 2 0 16:59 ? 00:00:00 [ksoftirqd/1]
only two tasklet thread are spawned to handle the these softirq at each core. i suspect some error in the interpreting the number of cores here, as there is only two core. (/proc/cpuinfo).
good reading on these internals:
http://lwn.net/Articles/308117/ http://www.wil.cx/matthew/lca2003/paper.pdf
-- Regards, Peter Teoh
-- Regards, Peter Teoh
participants (2)
-
Kernel Piddu -
Peter Teoh