Exact Extract — Huawei Training Material, Page 224 (Slide 36):
DPC supports NUMA affinity with compute applications, eliminating the latency of cross-NUMA memory access and improving the bandwidth performance of a single DPC client by over 30% in typical scenarios.
Huawei OceanStor Pacific Distributed Parallel Client (DPC) provides a high-performance data-access mechanism for distributed computing applications, particularly high-performance computing (HPC) workloads.
NUMA, or Non-Uniform Memory Access, is a multiprocessing architecture in which processors have faster access to their local memory than to memory attached to another processor or NUMA node.
When an application executes on one CPU while its storage-client processing threads access memory associated with another CPU, cross-NUMA memory transactions introduce additional latency.
Such remote-memory access can reduce application throughput and increase CPU overhead.
DPC addresses this problem through NUMA-affinity optimization.
Storage-client processing threads are associated with the NUMA resources used by their corresponding computing applications. This improves processor-cache locality and minimizes unnecessary cross-NUMA memory access.
For example, when an application runs on CPU 1 and writes data into CPU 1's local memory, DPC can process the relevant data using threads associated with the same NUMA domain.
The result is lower memory-access latency, reduced interconnect traffic, and improved storage-client bandwidth.
Huawei identifies bandwidth improvements exceeding 30% for a single DPC client under the specified typical test conditions.
Therefore, the statement is TRUE.