We investigated how to distribute computations among computational resources under data parallelism. A min–max model of computation times was proposed to reflect a heterogeneous computing system. Time functions for each resource were estimated with reference parameters, and sampling statistics evaluates those parameters such as effective memory bandwidths and FLOPS. Our min–max model includes those time functions as objective functions so that it suggests load balancing point for an arbitrary problem size. Several BLAS examples confirm that our model fits well comparing with real heterogeneous computing with OpenCL.