Re: [PATCH v2 01/12] sched/isolation: Use cases for fine-grained isolation

Re: [PATCH v2 01/12] sched/isolation: Use cases for fine-grained isolation

[Qiliang Yuan]

Hi Waiman,

On Mon, 13 Apr 2026 15:25:46 -0400, Waiman Long wrote:
> Usually, if we want to run a latency sensitive workload like DPDK, we 
> try to minimize all sorts of kernel noises or interference as much as 
> possible. Do you have a good use case where it is advantageous to remove 
> some types of kernel noises from a given set of CPUs but not the others?

The initial intent for fine-grained control was to allow RCU callback
processing to remain local on specific CPUs to mitigate grace-period
stalls on large NUMA systems.

However, I agree that for typical DPDK use cases, simplified full isolation
is preferred to reduce complexity. In v3, I will pivot to a unified
isolation state. The primary value we aim to provide is the ability to
transition CPUs into this state dynamically via cgroups, which eliminates
the need for GRUB-based reboots and allows us to deploy latency-critical
services without interrupting existing workloads.

Best regards,
Qiliang