Ampere Performance Toolkit
Ampere Performance Toolkit (APT)
The Ampere Performance Toolkit represents a significant advancement in simplifying software porting, benchmarking, and optimization. By streamlining the migration process, reducing development costs, and enabling access to a wider ecosystem, the toolkit empowers developers to embrace the benefits of the AArch64 architecture more quickly and effectively.
The Ampere System Profiler (ASP)
The Ampere System Profiler (ASP) provides profiling capabilities that can be used for workload characterization at a system level, a first step for root-causing performance issues. ASP can be used to study OS-level metrics, including CPU utilization, CPU Power, networking , I/O and NUMA effects, helping you resolve system-level bottlenecks before digging deeper into code and microarchitecture-level optimizations.
The Ampere Porting Advisor (APA)
The Ampere Porting Advisor (APA) centers around advising on the transition or optimization of software to Arm64 architecture. This tool can save time and effort in the porting and migration process. By analyzing the make environment and source code for known code patterns and dependency libraries it provides detailed reports with incompatibilities and recommendations. This guidance enables developers to navigate the intricacies of transitioning between architectures more efficiently, accelerating the overall migration process.
The Ampere PMU Profiler (APP)
The Ampere PMU Profiler (APP) focuses on deep, microarchitecture-level efficiency metrics by collecting Performance Monitoring Unit (PMU) events such as Instructions Per Cycle, Cache Miss Rate and Branch Prediction Analysis. Once system-level optimizations are completed, APP is critical to tying software code execution to hardware effects.
The Ampere Performance Kit for Benchmarking (APB)
The Ampere Performance Kit for Benchmarking (APB) provides a standardized and automated way to run and benchmark applications. This framework makes it easier and faster to set up and run performance tests in a repeatable manner across bare metal, virtualized, and cloud environments using a mature, automated framework with best known configurations. The APB project includes numerous benchmarks and workloads – from common benchmarks like multi-chase (memory latency/bandwidth), FIO (I/O characteristics), Netperf (network) and SPECCPU Integer to applications like Cassandra, MySQL, Redis, NGINX, and finally, AI inference using Large Language Models LLaMA and Qwen.