HP (Hewlett-Packard) HP IO Network Card User Manual


 
Performance and tuning 66
Setting NUMA affinity
Servers with a NUMA (Non-Uniform Memory Access) architecture require special installation instructions in
order to maximize ioMemory device performance. These servers include the HP ProLiant DL580 and HP
DL980 Servers.
On servers with NUMA architecture, during system boot, the BIOS on some systems will not distribute PCIe
slots evenly among the NUMA nodes. Each NUMA node contains multiple CPUs. This imbalanced
distribution means that, during high workloads, half or more of the CPUs might remain idle while the rest are
100% utilized. To prevent this imbalance, you must manually assign IO Accelerator devices equally among
the available NUMA nodes.
For information on setting NUMA affinity, see "NUMA configuration (on page 67)."
Setting the interrupt handler affinity
Device latency can be affected by placement of interrupts on NUMA systems. HP recommends placing
interrupts for a given device on the same NUMA socket that the application is issuing I/O from. If the CPUs
on this socket are overwhelmed with user application tasks, in some cases it might benefit performance to
move the interrupts to a remote socket to help load balance the system.
Many operating systems will attempt to dynamically place interrupts for you and generally make good
decisions. Hand tuning interrupt placement is an advanced option that requires profiling of application
performance on any given hardware. For information on how to pin interrupts for a given device to specific
CPUs, see your operating system documentation.