Resource allocation is a well-known problem, with a large number of research contributions towards efficient utilisation of the massive hardware parallelism using various exact and heuristic approaches.We address the problem of optimising resources usage on deeply heterogeneous platforms in the context of HPC systems running multiple applications with different quality of service levels. Our approach manages the partitioning within a single heterogeneous node aiming at serving as many critical applications as possible while leaving to the upper levels of runtime resource management the decision to preempt resources or to launch the critical application on a different node. We investigate predictive allocation algorithms, allowing to serve up to 20% more high priority requests when using a moving average or machine learning prediction model vs baseline without prediction.