Introduction to Kunpeng Precision Tuning
Precision is very important in the meteorological industry. Weather forecast is not only related to enterprise production and our daily life, but also related to disaster prevention. In bidding activities initiated by meteorological administrations, the bid inviting parties require that the forecast results of a new cluster be consistent with those of the existing cluster (generally x86) to ensure that the new cluster environment is accurate, stable, and efficient to be put into use. To ensure the accuracy and stability of the compute results, in the past, meteorological administrations had invested a large amount of workforce and gone through continuous tuning that last about six months when replacing cluster servers (including the replacement of IBM midrange servers with Intel x86 servers). Similarly, when a meteorological cluster is ported to the Kunpeng platform, the running result may be inconsistent with that of the existing cluster.
Besides the meteorological industry, for HPC applications, the results may be inconsistent when the cluster, CPU architecture, or compiler is replaced. Even a small difference can be amplified after a large number of iterations and calculations conducted by HPC. Therefore, the compute results differ greatly. Weather Research and Forecasting (WRF) model can help analyze and solve this problem. WRF is a model for fine-grained weather simulation and forecast, and is one of the important HPC application scenarios. For details, see Figure 1.
The indicators that customers are most concerned about in this model are listed in sequence of their priorities as follows:
- Accuracy: The difference between the forecast result and the real-time weather is small, such as the precipitation area and precipitation.
- Stability: The test results are consistent at different times, on different machines, and under different computing scales.
- Performance: Computing is fast to support timely weather forecast.
This document describes common precision tuning methods and analysis tools for Kunpeng HPC applications from four aspects: compiler, math library, MPI, and application.