NUMA Architecture: Why Isn't All Memory Equally Fast Inside a Server?

Introduction

When you purchase a server with two or more physical CPUs, it's easy to assume that all RAM can be accessed at the same speed. In reality, memory access latency depends on where the memory is physically located in relation to the processor. This concept is known as NUMA (Non-Uniform Memory Access) Architecture.

NUMA plays a significant role in the performance of databases, virtual machines, and high-performance applications running on multi-socket servers.

What Is NUMA Architecture?

NUMA (Non-Uniform Memory Access) is a memory architecture in which system memory is divided into multiple regions, each directly associated with a specific processor.

Every CPU can access all memory in the system, but it can access its own local memory much faster than memory attached to another processor.

How Does NUMA Work?

In a multi-processor server:

• Each processor has its own local memory.
• All processors can access every memory region.
• Accessing local memory is significantly faster than accessing remote memory connected to another CPU.
• This difference in access time is why it is called Non-Uniform Memory Access.

Why Is NUMA Used?

NUMA was designed to improve the scalability and efficiency of modern servers by:

• Enhancing performance on large multi-CPU systems.
• Reducing memory bus contention.
• Improving scalability as additional processors are added.

How Does NUMA Affect Performance?

If an application runs on one processor while frequently accessing memory attached to another processor:

• Memory latency increases.
• Overall application performance decreases.
• CPU utilization may rise due to longer wait times for memory access.

Applications Most Affected by NUMA

NUMA awareness is particularly important for:

• Database systems such as PostgreSQL and MySQL.
• Hypervisors like VMware and KVM.
• Kubernetes clusters.
• Artificial Intelligence and Machine Learning workloads.
• Big Data analytics platforms.

How Can You Optimize NUMA Performance?

To achieve the best performance on NUMA-enabled servers:

• Pin processes to specific CPUs using CPU Affinity.
• Allocate memory close to the processor executing the workload.
• Monitor workload distribution across NUMA nodes.
• Enable NUMA-aware settings whenever supported by the operating system or application.

FAQ

Does NUMA affect every server?

No. NUMA primarily impacts servers with two or more physical processors (multi-socket systems). Single-processor servers typically do not experience NUMA-related performance differences.

Can NUMA be ignored?

For lightweight workloads, NUMA may have little noticeable impact. However, for high-performance applications, databases, virtualization platforms, and AI workloads, ignoring NUMA can lead to measurable performance degradation.

Conclusion

Understanding NUMA Architecture is essential for maximizing the performance of multi-processor servers. By ensuring that applications access memory located close to the processor executing them, organizations can reduce memory latency, improve resource utilization, and achieve better overall system performance.