Concept and Mechanism of MPAM

MPAM is a technology used in the AArch64 architecture to solve performance problems caused by resource sharing (memory bandwidth and L3 cache) on a server.

MPAM copes with the problem that the performance of some key applications or the overall system performance deteriorates due to resource contention (such as the cache, DMC, or interconnect) when different types of services are deployed on a server. It adds different labels (such as PARTID and PMG) to all requests of different service flows in the CPU or system memory management units (SMMUs) so that the hardware can detect the service flows. Based on the label information, system resources (such as the cache and DMC bandwidth) can be dynamically allocated to isolate service flows from each other and reduce mutual interference.

After MPAM is enabled for Kubernetes, you can specify a resource group when creating a Pod. As a result, resources accessed by an application are isolated from those accessed by other applications. In this way, resource usage is monitored and resource utilization is improved. In addition, you can analyze the monitoring data to locate the causes of performance deterioration of key applications and take countermeasures in a timely manner.