Introduction

The best practices described in this document demonstrate the functions of each tool of the Kunpeng DevKit. You can refer to these best practices when analyzing and optimizing your development projects in the Kunpeng DevKit. See Table 1.

**Table 1** Best practices
Tool	Feature	Working Mode	Best Practice	Description
System Profiler Kunpeng Performance Boundary Analyzer	Microarchitecture analysis	CLI	Practice 1: Microarchitecture Analysis	In this practice, the Kunpeng Performance Boundary Analyzer is used to quickly identify performance issues. Preliminary analysis indicates a high branch misprediction rate among microarchitecture metrics, suggesting the presence of a performance bottleneck. The System Profiler is then used to analyze the microarchitecture. The CPU branch misprediction rate for conditional statements is found to be high. Examination of the source code reveals that data is not processed before these conditional statements are executed. To address this problem, the data is sorted in the source code, optimizing CPU branch prediction, increasing the branch prediction success rate, and improving overall application performance.
	Hotspot function analysis	CLI	Practice 2: Hotspot Function Analysis	In this practice, the Kunpeng Performance Boundary Analyzer is used to quickly identify performance issues. It is preliminarily identified that hotspot system functions are frequently invoked, indicating a performance bottleneck. Then, the System Profiler is used to analyze the hotspot functions, and a flame graph is used to examine the call stack. The analysis shows that I/O system calls account for a high proportion of execution time. Source code inspection further indicates that this issue is caused by high overhead from frequent read system calls. Next, the memory mapping (mmap) method is used to reduce data copies and system calls, optimizing the large-file read logic and thereby reducing I/O latency and improving program performance.
	Memory access statistics analysis	CLI	Practice 3: Memory Access Statistics Analysis	In this practice, the Kunpeng Performance Boundary Analyzer is used to quickly identify performance issues. Preliminary analysis indicates that DDRC read bandwidth is extremely high, suggesting a performance bottleneck. The System Profiler is then used to examine memory access statistics and cache miss events, revealing a low cache hit ratio in the application. Further analysis of the source code shows that this issue is caused by extensive memory data replication. Block processing is applied to increase the cache hit ratio and enhance program performance.
System Profiler	Hotspot function analysis	WebUI	Practice 1: Tuning Python String Concatenation	Use the System Profiler to perform hotspot function analysis on Python string concatenation APIs, locate performance bottlenecks, and tune the performance of Python string concatenation based on the analysis result.
System Profiler	Hotspot function analysis Lock and wait analysis	WebUI	Practice 2: Tuning Lock Performance	Use the hotspot function analysis and lock and wait analysis functions of the System Profiler to sample and analyze multi-thread applications in the target environment, locate performance bottlenecks, and tune the application performance based on the analysis result.
Java Profiler	Real-time profiling Sampling profiling	WebUI	Practice 1: Tuning Memory Leaks	Use the Java Profiler to perform real-time profiling and sampling profiling on a running Java program, locate program problems, and tune the program based on the analysis results to achieve optimal running of the Java program.
	Real-time profiling Sampling profiling	WebUI	Practice 2: Locating and Tuning Hotspot Functions	Use the Java Profiler to perform hotspot function analysis on a running Java program, locate program hotspot problems, and tune the program based on the analysis results to achieve optimal running of the Java program.
	Real-time profiling	WebUI	Practice 3: Tuning GC Logs	Use the Java Profiler to perform garbage collection (GC) analysis on a running Java program, locate program GC problems, and tune the program based on the analysis results to achieve optimal running of the Java program.
		WebUI	Practice 4: Locating and Tuning Lock Contention	Use the Java Profiler to perform thread dump (with CPU analysis) on a running Java program, locate program lock problems, and tune the program based on the analysis results to achieve optimal running of the Java program.
		WebUI	Practice 5: Tuning GC Triggered by G1 Humongous Objects	Use the Java Profiler to perform GC analysis on a running Java program, locate program GC problems, and tune the program based on the analysis results to achieve optimal running of the Java program.
		WebUI	Practice 6: Tuning Abnormal GC	Use the Java Profiler to perform GC analysis on a running Java program, locate program GC problems, and tune the program based on the analysis results to achieve optimal running of the Java program.
		WebUI	Practice 7: Tuning Deadlocks	Use the Java Profiler to perform thread dump (with CPU analysis) on a running Java program, locate program lock problems, and tune the program based on the analysis results to achieve optimal running of the Java program.
System Diagnosis	Memory usage	WebUI	Practice 1: Tuning Memory Usage	Use the Kunpeng DevKit System Diagnosis tool to analyze the memory usage of executable programs, locate memory leaks based on the call stack information, and optimize memory usage accordingly.

Parent topic: Best Practices