Dynamic phases borrow from dynamic timers and static phases to create performance data for all routines that are invoked in a given invocation of a phase. If we instrument a routine as a dynamic phase, creating a unique name for each of its invocations (by embedding the invocation count in the name), we can examine the time spent in all routines and child phases invoked directly or indirectly from the given phase. The performance data generated can typically answer questions such as: what is the total time spent in MPI_Send() when it was invoked directly or indirectly in iteration 24? Dynamic phases are useful for tracking per-iteration profiles for an adaptive computation where iterations may differ in their execution times.