Saxpy Performance

Ripple has been benchmarked for the simple saxpy example against other libraries to evaluate the overhead of the implementation which facilitates the expressive api. Specifically, we benchmark against cuBLAS and Kokkos, since cuBLAS is the GPU library for linear algebra, and Kokkos also simplifies GPU programming.

To see the actual code implementations, see benchmarks/saxpy from the root directory of ripple. The results presented here were run on a V100 GPU. All times are in milliseconds.

Performance comparison for saxpy

Elements	cuBLAS	Kokkos	Ripple	Ripple NBC
1M	0.0749	0.0421	0.0327	0.0291
10M	0.2316	0.2722	0.1704	0.1119
100M	1.5286	2.5773	1.5242	1.4622
200M	2.9785	5.1334	3.0356	2.9088
500M	7.3153	12.826	7.5261	7.2347
1B	14.556	25.618	15.048	14.439

Note

Ripple NBC is the performance using ripple after removing the check at the end of the domain to determine if the cell is inside the domain specified by the tensor. While this is required, we add the result to illustrate that the cost of the check is minimal, being a few percent for even the simplest kernel.

So we see here that the performance of ripple is very good, even for something as simple as saxpy, which is not what ripple is designed for, which is for more complex computations across multiple gpus.