Configuring CPU Affinity

Introduction

Process-to-core binding (also known as CPU affinity) is when you bind processes or threads to specific CPU cores. This way, a process can only be scheduled to specific cores and process locations can be controlled more precisely.

Preparations

Prepare an ECS.

Prerequisites

• Application scenarios: Frequent cross-core migration, or cross-NUMA or cross-socket migration impacts performance.
• Effects: Each process is bound to specific CPU cores to avoid cross-core, cross-NUMA, or cross-die migration.

Procedure

1. Bind a process to CPU cores.

   Bind a process to CPU cores when you start it.

   taskset -c {cpu_lists} process

   Bind a running process to CPU cores.

   taskset -cp {cpu_lists} PID

2. If NUMA is used, you can also use a numactl command to bind a process to CPU cores.

   Bind a process to CPU cores.

   numactl -C {cpu_list} process
   #-C cpus or --physcpubind=cpus

   Bind a process to NUMA nodes.

   numactl -N {numa_node_list} process
   # -N nodes or --cpunodebind=nodes

• {cpu_list}: for example, 0, 1 to 13, or 16
• {numa_node_list}: for example, 0, 0 and 2, or 0 to 2
• To check the CPU affinity of a process, run taskset -cp {pid}.
• Other numactl parameters:
  • -m {numa_node} binds memory. For example, -m 0 binds memory to numa0. -m 0,1 binds memory to numa0 and numa1. -m 0-2 binds memory to numa0, numa1, and numa2.
  • -i <nodes> controls interleaved memory allocation. For example, -i 0,1 configures memory to be interleaved across numa0 and numa1. -i all configures memory to be interleaved across all NUMA nodes.