306
Understanding the NUMA behavior of your workload is another important step in
understanding performance optimization. Figure 15-8 shows UPI bandwidth collected
with VTune Profiler.
The Platform Profiler feature in VTune Profiler can collect metrics specific to
persistent memory.
Tuning the Hardware
The memory configuration of a system is a significant factor in determining the system’s
performance. The workload performance depends on a combination of workload
characteristics and the memory configuration. There is no single configuration that
provides the best value for all workloads. These factors make it important to tune the
hardware with respect to workload characteristics and get the maximum value out of the
system.
Addressable Memory Capacity
The combined capacity of DRAM and persistent memory determines the total
addressable memory available on the system. You should tune the size of persistent
memory to accommodate the workload’s footprint.
The capacity of DRAM available on the system should be large enough to
accommodate the workload’s hot working set size. A large amount of volatile traffic going
to persistent memory while DRAM is fully utilized is a good indicator that the workload
can benefit from additional DRAM size.
Figure 15-8. UPI traffic ratio from VTune Profiler
Chapter 15 profiling and performanCe
Understanding the NUMA behavior of your workload is another important step in
understanding performance optimization. Figure 15-8 shows UPI bandwidth collected
with VTune Profiler.
The Platform Profiler feature in VTune Profiler can collect metrics specific to
persistent memory.
Tuning the Hardware
The memory configuration of a system is a significant factor in determining the system’s
performance. The workload performance depends on a combination of workload
characteristics and the memory configuration. There is no single configuration that
provides the best value for all workloads. These factors make it important to tune the
hardware with respect to workload characteristics and get the maximum value out of the
system.
Addressable Memory Capacity
The combined capacity of DRAM and persistent memory determines the total
addressable memory available on the system. You should tune the size of persistent
memory to accommodate the workload’s footprint.
The capacity of DRAM available on the system should be large enough to
accommodate the workload’s hot working set size. A large amount of volatile traffic going
to persistent memory while DRAM is fully utilized is a good indicator that the workload
can benefit from additional DRAM size.
Figure 15-8. UPI traffic ratio from VTune Profiler
Chapter 15 profiling and performanCe