Computing performance metrics and memory bandwidth bottlenecks

Technical analysis highlights the importance of measuring resource contention and data movement in computing performance. Linux Pressure Stall Information (PSI) provides kernel metrics to track how much time workloads spend stalled due to CPU, memory, or I/O contention. Unlike traditional utilization metrics, which only show how much a resource is being used, PSI identifies whether contention is actively preventing work from progressing. On Linux systems, these measurements are exposed via /proc/pressure/ files, allowing administrators to distinguish between host-wide pressure and pressure affecting specific process groups via cgroup v2.

In High Performance Computing (HPC), memory bandwidth serves as a critical factor in scaling performance. While memory capacity determines how much data can be stored, bandwidth dictates the speed at which data moves between memory and the processor. A memory bandwidth bottleneck occurs when a CPU can process data faster than the memory subsystem can supply it, causing cores to sit idle. This limitation often prevents HPC applications from achieving linear performance scaling when increasing the number of CPU cores, as additional cores increase the demand for data without a proportional increase in delivery speed.