Chip Alignment in LPAR

Edit online

Best Practices to Optimize AI Services Performance

To achieve maximum efficiency and performance for AI workloads, follow these best practices when configuring LPARs and system resources:

  1. Contain the workload within a single chip
    • Create an LPAR that fits entirely on one chip for optimal performance.
  2. Keep affinity score inside the LPAR in check

    A score of 100 indicates that CPU and memory resources are optimally allocated. Lower scores reflect reduced resource locality, which may negatively impact performance.

    Run the following command to discover the affinity score:

    cat /proc/ppc64/lparcfg | grep affinity

    Learn more about Affinity Score, here.

  3. Follow IBM Power Virtualization Best Practices

    We recommend adhering to the IBM Power Virtualization best practices outlined here to ensure optimal performance and reliability.

  4. Use Dedicated Processors

    We recommend using dedicated processors to ensure consistent performance, reduce resource contention, and improve workload isolation.

  5. Use Dynamic Platform Optimization for NUMA-Related Performance Issues

    Dynamic Platform Optimization (DPO) shall be utilized when performance concerns can be attributed to NUMA effects.