Linux non-Uniform memory distribution

Linux non-Uniform memory distribution

Directory[show]Traditional SMP Architecture

NUMA, non-uniform memory access (Non-uniform memories Access), between SMP and MPP. In the NUMA architecture, each CPU is called a node, and the memory between each node is used independently.

Each CPU is absolutely equal, there is no priority, access to memory must pass through the system bus. Access to the CPU also needs to go through the system bus.

From this architecture, you can also see where the short board of the SMP architecture is. This is obviously problematic for systems that are now prone to dozens of or even hundreds of CPUs. The system's bus will be the bottleneck of the whole system.

NUMA Architecture

With the development of technology, a new kind of architecture Numa is introduced. Each CPU is independent of each other and the memory between them is not affected. Each CPU accesses its own memory, and the latency is minimal.

Numa Configuration Method

1) Turn off NUMA settings in the BIOS

2) Turn off NUMA at the operating system kernel level, for example:

/etc/grub.conf's kernel line was last added: Numa=off

3) application shutdown such as Oracle database plane shutdown:

_enable_numa_optimization=false (parameter in 11g is _enable_numa_support)

Numa common Commands

To view NUMA statistics:

Numastat

View CPU Attribution node

Lscpu

The Numactl--hardware command returns the total memory size of the different nodes, the available size, and the node distance information

Linux non-Uniform memory distribution