Pinning nvme interrupts to a single cpu?

Pinning nvme interrupts to a single cpu?

[Jeff Johnson]

Greetings,

Is it possible to limit nvme interrupts to a single cpu, rather than
spread them across many? I have a latency sensitive workload and I am
trying to do the opposite of what many appear to be doing, I need to
assign interrupts to a small number of cpus (one for nvme, one for
Ethernet, etc) leaving the bulk of my cpus for the application.

Single nvme drive which is both boot and application local storage.
Performance is nice but in my case it's more important to pin the
interrupts to one cpu even if local nvme performance decreases.

Thx,
--Jeff

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: Pinning nvme interrupts to a single cpu?

[Keith Busch]

On Mon, Apr 27, 2020 at 05:10:47PM -0700, Jeff Johnson wrote:
> Is it possible to limit nvme interrupts to a single cpu, rather than
> spread them across many?

The managed interrupts provided by the irq subsystem that the nvme
driver uses doesn't allow altering the cpu affinity. The only way to
have all interrupts handled by a single cpu is to have only one nvme IO
queue, but you don't get to choose the cpu.

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: Pinning nvme interrupts to a single cpu?

[Ming Lei]

On Tue, Apr 28, 2020 at 8:11 AM Jeff Johnson
<jeff.johnson@aeoncomputing.com> wrote:
>
> Greetings,
>
> Is it possible to limit nvme interrupts to a single cpu, rather than
> spread them across many? I have a latency sensitive workload and I am
> trying to do the opposite of what many appear to be doing, I need to
> assign interrupts to a small number of cpus (one for nvme, one for
> Ethernet, etc) leaving the bulk of my cpus for the application.
>
> Single nvme drive which is both boot and application local storage.
> Performance is nice but in my case it's more important to pin the
> interrupts to one cpu even if local nvme performance decreases.

You might reach that by passing isolcpus.managed_irq to kernel for
isolating CPUs which needn't to handle NVMe IO.

thanks,
Ming Lei

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme