A partitioned NUMA machine is managed to dynamically transform its partition layout state based on NUMA considerations. The NUMA machine includes two or more NUMA nodes that are operatively interconnected by one or more internodal communication links. Each node includes one or more CPUs and associated memory circuitry. Two or more logical partitions each comprise at a CPU and memory circuit allocation on at least one NUMA node. Each partition respectively runs at least one associated data processing application. The partitions are dynamically managed at runtime to transform the distributed data processing machine from a first partition layout state to a second partition layout state that is optimized for the data processing applications according to whether a given partition will most efficiently execute within a single NUMA node or by spanning across a node boundary. The optimization is based on access latency and bandwidth in the NUMA machine.