How to use the command 'numactl' (with examples)

The numactl command allows users to control Non-Uniform Memory Access (NUMA) policy for processes or shared memory in systems equipped with multiple processors. NUMA is an architecture that allocates memory blocks to specific processors, making the execution of processes more efficient by reducing latency. By using numactl, system administrators and power users can finely tune performance by specifying on which CPU or memory node processes should run. This capability enhances performance optimization, especially in high-performance computing applications that demand efficient resource management.

Use case 1: Run a command on node 0 with memory allocated on node 0 and 1

Code:

numactl --cpunodebind=0 --membind=0,1 -- command command_arguments

Motivation: In a NUMA system, different memory nodes can exhibit varying performance characteristics based on their proximity to a processor node. By binding a command to run solely on CPUs of node 0 while allowing memory allocations from nodes 0 and 1, it is possible to minimize memory latency. This approach can lead to significant performance improvements in data-intensive applications where memory access speed is crucial.

Explanation:

• --cpunodebind=0: This argument confines the command to execute exclusively on the CPUs associated with node 0. It restricts the process from using CPUs from any other nodes, ensuring that all CPU operations occur on node 0.
• --membind=0,1: This binds the memory allocation to nodes 0 and 1. The command is allowed to allocate memory from these nodes only, which can help balance memory load and optimize access times.
• command command_arguments: This is the command and its arguments that you wish to execute with the specified NUMA constraints.

Example Output: When implemented in practice, you might observe lower memory latency for memory-intensive applications, potentially resulting in faster application processing times. There won’t always be immediate visual output from this command, but performance monitoring tools can highlight improved CPU and memory efficiency.

Use case 2: Run a command on CPUs (cores) 0-4 and 8-12 of the current cpuset

Code:

numactl --physcpubind=+0-4,8-12 -- command command_arguments

Motivation: Selecting specific CPU cores for command execution can lead to optimized CPU usage and reduced context-switching overhead. This use case is especially beneficial for multi-threaded applications that require dedicated CPU cores to maximize throughput and minimize interruptions from other processes.

Explanation:

• --physcpubind=+0-4,8-12: This argument ensures that the command runs solely on the specified CPU cores: 0 through 4 and 8 through 12 within the current cpuset. By binding the operation to these cores, you can harness the computational power of selected CPUs while preventing interference from other cores.
• command command_arguments: This segment is the user-defined command and its associated parameters that you wish to execute.

Example Output: By constraining a command to specified CPU cores, more predictable CPU interruption behavior is achieved, which helps ensure that performance is consistent and aligns closely with application requirements. Monitoring tools may show concentrated CPU utilization on the chosen cores, with reduced impact on adjacent processes.

Use case 3: Run a command with its memory interleaved on all CPUs

Code:

numactl --interleave=all -- command command_arguments

Motivation: Using memory interleaving across all CPU nodes is a potent method for achieving a balanced memory load across the system. In scenarios where memory access rates are critical, interleaving can help reduce contention by distributing memory requests evenly. This creates a symmetrical access pattern, which is beneficial for workloads that involve consistent memory usage across multiple processors.

Explanation:

• --interleave=all: Interleaving memory across all nodes ensures that memory access is balanced, reducing potential contention points. This is particularly beneficial for applications that are sensitive to memory bandwidth and latency.
• command command_arguments: This is the executable command and its parameters that will be executed with the interleaved memory configuration.

Example Output: The actual performance gain will depend on the specific application and its memory usage pattern, but the goal is to achieve a more even distribution of memory access, which reduces variance in access times. Monitoring tools would indicate a uniform memory utilization across nodes, illustrating the distribution achieved through interleaving.

Conclusion:

Understanding and utilizing the numactl command allows for strategic resource allocation, maximizing CPU and memory performance in a NUMA system. Each of the illustrated use cases provides a method to optimize applications by adjusting how they leverage system resources. By mastering these techniques, users can significantly enhance application performance, reduce latency, and improve overall system efficiency.