Dynamic allocation of cloud computing processors

An efficient architecture and method is provided for implementing improvement or optimization of cloud computing processing resources. For machine learning computations including parallel processing, these techniques may include reallocating recently available cloud processing resources based on parallelization metrics. Such improved resource allocation may use a first cost function to improve or optimize processing resource utilization, and may use a second cost function to reduce or minimize computational time. These cost functions may be solved in parallel using a graph model to characterize nodes and edges to improve or optimize dynamic reallocation of cloud computing processing resources. This can be used to improve the overall cloud computing performance of any processor, and can enable specific improvements in convolution calculations performed in a graphics processing unit (GPU).