303
• The writes to persistent memory may impact performance more than
the reads.
• Applications can allocate objects on DRAM or persistent memory.
If done indiscriminately, this can negatively impact performance.
• In Memory Mode (specific to Intel Optane DC persistent memory),
users have the option of varying the near-memory cache size (DRAM
size) to improve workload performance.
Keeping these additional factors in mind, the approach to workload performance
optimization will follow the same process of characterizing the workload, choosing the
correct memory configuration, and optimizing the code for maximum performance.
Characterizing the Workload
The performance of a workload on a persistent memory system depends on a
combination of the workload characteristics and the underlying hardware. The key
metrics to understand the workload characteristics are:
• Persistent memory bandwidth
• Persistent memory read/write ratio
• Paging to and from traditional storage
• Working set size and footprint of the workload
• Nonuniform Memory Architecture (NUMA) characteristics
• Near-memory cache behavior in Memory Mode (specific to Intel
Optane DC persistent memory)
Memory Bandwidth and Latency
Persistent memory, like DRAM, has limited bandwidth. When it becomes saturated, it
can quickly bottleneck application performance. Bandwidth limits will vary depending
on the platform. You can calculate the peak bandwidth of your platform using hardware
specifications or a memory benchmarking application.
Chapter 15 profiling and performanCe
• The writes to persistent memory may impact performance more than
the reads.
• Applications can allocate objects on DRAM or persistent memory.
If done indiscriminately, this can negatively impact performance.
• In Memory Mode (specific to Intel Optane DC persistent memory),
users have the option of varying the near-memory cache size (DRAM
size) to improve workload performance.
Keeping these additional factors in mind, the approach to workload performance
optimization will follow the same process of characterizing the workload, choosing the
correct memory configuration, and optimizing the code for maximum performance.
Characterizing the Workload
The performance of a workload on a persistent memory system depends on a
combination of the workload characteristics and the underlying hardware. The key
metrics to understand the workload characteristics are:
• Persistent memory bandwidth
• Persistent memory read/write ratio
• Paging to and from traditional storage
• Working set size and footprint of the workload
• Nonuniform Memory Architecture (NUMA) characteristics
• Near-memory cache behavior in Memory Mode (specific to Intel
Optane DC persistent memory)
Memory Bandwidth and Latency
Persistent memory, like DRAM, has limited bandwidth. When it becomes saturated, it
can quickly bottleneck application performance. Bandwidth limits will vary depending
on the platform. You can calculate the peak bandwidth of your platform using hardware
specifications or a memory benchmarking application.
Chapter 15 profiling and performanCe