Multi-core nodes with Non-Uniform Memory Access (NUMA) are now a common architecture for high performance computing. On such NUMA nodes, the shared memory is physically distributed into memory banks connected by a network. Owing to this, memory access costs may vary depending on the distance between the processing unit and the memory bank. Therefore, a key element in improving the performance on these machines is dealing with memory affinity. We propose a NUMA-aware load balancer that combines the information about the NUMA topology with the statistics captured by the Charm++ runtime system. We present speedups of up to 1.8 for synthetic benchmarks running on different NUMA platforms. We also show improvements over existing load balancing strategies both in benchmark performance and in the time for load balancing. In addition, by avoiding unnecessary migrations, our algorithm incurs up to seven times smaller overheads in migration, than the other strategies.