Kunpeng Affinity Optimization
The Kunpeng BoostKit for Big Data algorithm library provides optimizations to greatly improve the computing performance in big data algorithm scenarios.
Memory Access Optimization
Some algorithms, such as PCA, require calculating the Gram matrix. The Kunpeng BoostKit algorithm library uses the large capacity of Kunpeng to partition and rearrange the matrix, increasing the cache hit ratio. It improves the performance of algorithms such as PCA by more than 50% at the same accuracy level.
Figure 1 Kunpeng BoostKit solution
Multi-core Parallel Computing
In tree model algorithms, data parallelism and model parallelism are combined to improve the parallelism degree of the algorithms while maintaining the communication volume, thereby giving full play to the multi-core advantage of Kunpeng. The Kunpeng BoostKit algorithm library improves the computing performance by more than 50% for tree model algorithms such as Random Forest and GBDT.
Figure 2 Kunpeng BoostKit solution
Parent topic: Key Technologies