On BG/Q, we have only one type of link to test. We measure the ping-pong latency between two directly connected nodes. Dif-ferently from P7-IH, BG/Q hardware provides hardware reliability. Figure 9 compares the latency of our runtime and MPI_Send with different message sizes. On this network our runtime significantly 2L2 Atomics allow a single load or store to perform a simple arith-metic operation in the L2 cache with a much lower overhead than traditional atomics.