What performance considerations should I be aware of when running Red Hat Enterprise Linux 8 on IBM Power10 systems?
Environment
- IBM Power10 Systems
- Red Hat Enterprise Linux 8 for Power, little endian
Issue
In rare instances, a surplus of system resources on an IBM Power10 CPU can result in one core and its associated simultaneous multithreading (SMT) threads being observably more performant than other cores on the same CPU when accessing main memory.
Resolution
This is expected behavior and should only be seen in high-cache-miss scenarios or when workloads are running in a pool with an odd number of cores.
Root Cause
This phenomenon is an effect of the way the cores on an IBM Power10 CPU are distributed among multiple resource pools. When each core in the pool requests the maximum memory bandwidth for a task, the system provides a fair share of bandwidth equal to the percentage that the core represents in the pool. For example, if there are two cores in a pool, each core would receive 50% of the resources. In the event that one of the cores in that pool is being utilized at less than 50%, the remainder of the resources would be available to the other core, allowing it to perform at a higher level than cores in pools where all member cores are requesting full and equal utilization.
There are several features that can be affected by this behavior, including kernel shared memory (KSM), transparent hugepages (THP), and autoNUMA as they are memory intensive scanners.
This solution is part of Red Hat’s fast-track publication program, providing a huge library of solutions that Red Hat engineers have created while supporting our customers. To give you the knowledge you need the instant it becomes available, these articles may be presented in a raw and unedited form.
Comments