In the current implementation, there are two limitations that we do not fully address. First, we do not consider direct memory access (DMA) by other hardware components on nodes that may be in reduced power states, which may result in performance degradation. This can be mitigated by ensuring that DMA uses only pages within a pre-defined physical memory range (e.g., the first node), which, due to the use of library aggregation, is almost always in Standby mode.4
Second, kernel threads that run in the background may touch random pages belonging to any process in the system. Since these maintenance threads are invoked fairly infrequently, a simple solution is to treat these as special processes and turn on all nodes when they are invoked to avoid performance degradation.