Adaptive CPU NUMA scheduling

A method for assigning a process to one of a set of non-uniform memory access (NUMA) nodes comprising a first NUMA node and a second NUMA node, first NUMA node and the second NUMA node each including a destination central processing unit (CPU) and a memory, the method comprising:

• determining a working set size associated with the process;
• for a first destination CPU on the first NUMA node,
• determining a first intra-node memory access latency for the first destination CPU accessing a first memory on the first NUMA node;
• determining a first inter-node memory access latency for the first destination CPU accessing a second memory on the second NUMA node, the second NUMA node being remote from the first NUMA node;
• and
• determining a first memory access cost value for the first destination CPU on the first NUMA node based on the first intra-node memory access latency, the first inter-node memory access latency, and the working set size;
• for a second destination CPU on the second NUMA node,
• determining a second intra-node memory access latency for the second destination CPU accessing the second memory on the second NUMA node;
• determining a second inter-node memory access latency for the second destination CPU accessing the first memory on the first NUMA node;
• and
• determining a second memory access cost value for the second destination CPU on the second NUMA node based on the second intra-node memory access latency, the second inter-node memory access latency, and the working set size;
• accessing a first CPU cost value for the first destination CPU and a second CPU cost value for the second destination CPU;
• determining a migration suitability metric based on the first memory access cost value, the second memory access cost value, the first CPU cost value and the second CPU cost value;
• selecting an optimal destination CPU for placement of the process, the optimal destination CPU being selected from the first destination CPU and the second destination CPU based at least in part on the migration suitability metric;
• assigning the process to the selected optimal destination CPU;
• and
• increasing a scheduling priority of the process based on the first memory access cost value for the first destination CPU and the second memory access cost value for the second destination CPU to compensate for inefficient execution due to remote memory accesses.

One or more non-transitory computer-readable storage media including computer-executable components for assigning a process to one of a set of non-uniform memory access (NUMA) nodes comprising a first NUMA node and a second NUMA node, the first NUMA node and the second NUMA node each including a destination central processing unit (CPU) and a memory, the computer-executable components comprising:

• a memory cost estimator configured to cause the at least one processor to:
• determine a working set size associated with the process;
• for a first destination CPU on the first NUMA node,
• determine a first intra-node memory access latency for the first destination CPU accessing a first memory on the first NUMA node;
• determine a first inter-node memory access latency for the first destination CPU accessing a second memory on the second NUMA node, the second NUMA node being remote from the first NUMA node;
• and
• determine a first memory access cost value for the first destination CPU based on the first intra-node memory access latency, the first inter-node memory access latency, and the working set size;
• and
• for a second destination CPU on the second NUMA node,
• determine a second intra-node memory access latency for the second destination CPU accessing the second memory on the second NUMA node;
• determine a second inter-node memory access latency for the second destination CPU accessing the first memory on the first NUMA node;
• and
• determine a second memory access cost value for the second destination CPU on the second NUMA node based on the second intra-node memory access latency, the second inter-node memory access latency, and the working set size;
• a CPU cost estimator configured to cause the at least one processor to access a first CPU cost value of the first destination CPU and a second CPU cost value of the second destination CPU;
• and
• a CPU scheduler configured to cause the at least one processor to at least:
• determine a migration suitability metric based on the first memory access cost value, the second memory access cost value, first CPU cost value and the second CPU cost value, and wherein selecting the optimal destination CPU is based on the migration suitability metric;
• select an optimal destination CPU for placement of the process, the optimal destination CPU being selected from the first destination CPU and the second destination CPU based at least in part on the migration suitability metric;
• assign the process to the selected optimal destination CPU;
• and
• increase a scheduling priority of the process based on the first memory access cost value for the first destination CPU and the second memory access cost value for the second destination CPU to compensate for inefficient execution due to remote memory accesses.