How to Supercharge Kunpeng CPUs: Real‑World Performance Tuning Techniques

Why Performance Tuning Matters: A 4800×4800 Matrix Multiplication Example

Implementing a 4800×4800 matrix multiplication in Python takes about 61,162 seconds, while a single‑threaded C version reduces it to 757 seconds. Parallel C code brings the time down to 47 seconds, and further cache‑aware parallelism cuts it to 6.02 seconds. Using Kunpeng’s NEON vector instructions finally achieves 1.99 seconds, demonstrating the dramatic impact of hardware‑aware optimization.

Performance Tuning from a Von Neumann Perspective

The Von Neumann model abstracts a computer into CPU/memory, network, storage, and application layers. Optimizing each layer—CPU/memory, NIC, disk, and the application—forms the basis of systematic performance improvement.

Kunpeng Soft and Hard Acceleration Overview

Kunpeng CPUs provide both software acceleration (single‑core and multi‑core) and hardware acceleration (chip‑level engines) for workloads such as compression, encryption, and multimedia.

Compiler and JDK Optimizations for Single‑Core Speedup

Huawei’s compiler applies several techniques:

Instruction layout optimization: split functions and reorder hot/cold code to improve instruction‑cache hit rate.

Memory layout optimization: group frequently accessed data to boost data‑cache hit rate.

Loop optimization: parallelize independent loops across cores and auto‑vectorize dependency‑free loops.

For Java developers, the Bi‑Sheng JDK adds:

JIT compilation and GC improvements for faster memory management.

JVM loop, vector, and serialization enhancements to increase execution speed.

NUMA‑Based Multi‑Core Optimization

Traditional SMP architectures share a single memory controller, creating a bottleneck as core counts rise. NUMA partitions cores into nodes, each with its own memory controller, reducing contention. By binding processes to local NUMA nodes (e.g., using numactl -C 0-15 <process> or sched_setaffinity), memory access latency is minimized.

NUMA Optimization in Nginx

Binding the NIC, kernel network stack, and Nginx workers to the same NUMA node reduces double‑copy overhead, yielding roughly a 15 % reduction in end‑to‑end latency.

Building Acceleration Libraries on Kunpeng

Kunpeng offers nine acceleration libraries covering basic, compression, cryptography, and multimedia workloads, delivering 10‑100 % performance gains. For example, OpenSSL can offload cryptographic operations to a hardware engine without code changes, and Kunpeng’s high‑efficiency compression engines dramatically shorten gzip/zlib/ZSTD/snappy compression times.

Case Study: JD.com RSA Acceleration

Switching JD.com’s web services from a traditional QAT card to Kunpeng’s RSA acceleration engine increased HTTPS short‑connection performance by 33 %.

Disk and NIC Optimizations for a Better Runtime Environment

Choosing XFS, enabling file prefetch, tuning dirty‑page flushing, and applying I/O scheduler tweaks improve disk‑to‑memory throughput. Adjusting NIC interrupt rates, enabling interrupt coalescing, and configuring TSO/CSUM balance latency and throughput.

Application‑Level Tuning Strategies

Increasing concurrency, leveraging data caches, and using asynchronous I/O can boost performance, but excessive threads or cache size may cause lock contention and cache‑line thrashing. Techniques such as lock‑free programming, reducing lock granularity, and using high‑performance atomic instructions help. Tcmalloc reduces allocation‑related locks, and aligning data structures to Kunpeng’s 128‑byte cache line avoids false sharing.

MariaDB Performance‑Tuning Workflow

The process consists of three iterative phases: monitoring, analysis, and optimization.

Monitoring: collect CPU, memory, disk, and NIC metrics (e.g., interrupts, NUMA hit rate, iowait, bandwidth).

Analysis: identify bottlenecks such as CPU lock contention using perf or high TPS drop.

Optimization: adjust parameters like innodb_thread_concurrency, innodb_sync_spin_loops, and innodb_spin_wait_delay to improve throughput.

Kunpeng Performance‑Tuning “Ten‑Hammer” Checklist

Key tuning knobs:

CPU/Memory: adjust page size, enable CPU prefetch, modify thread scheduling policy.

Disk: tune dirty‑page flushing, use asynchronous I/O (libaio), adjust filesystem parameters.

NIC: enable multi‑queue, turn on TSO, enable checksum offload.

Application: optimize compile options, implement file‑cache mechanisms, cache execution results, and leverage NENO instruction acceleration.

Conclusion

CPU/memory, disk, NIC, and application are the four main dimensions of performance tuning.

Collect metrics, analyze bottlenecks, and iteratively optimize code and configuration.

Fully exploiting hardware resources is essential for achieving optimal software performance.

Balancing latency, throughput, and concurrency is crucial for stable, high‑performance services.