We start with a ping-pong analysis to understand the latency across different links. On P7-IH, the network supports three dif-ferent physical links: LL, LR and D [18]. Different links are used to connect different levels of the network. We initially focus on the link latency, comparing our runtime to MPI_Send. The P7-IH network does not implement a reliable hardware packet transport layer. With our runtime we can choose between a thread-level flow control implementation or the native, unreliable connection. MPI always uses software-based flow control.