A simple ping-pong benchmark using am_request_1 and am_reply_1 shows a one-word round-trip latency of 51.0 s on thin nodes. This value increases by about 0.25 s per word when two, three, or four 32-bit words are transferred. This round trip latency compares well with a raw message (no data or sequence number) ping-pong latency of 46.6-47.0 s. The additional overhead of 4 s is due to the cost of the cache flushes and the flow control bookkeeping. The same ping-pong test using MPL's mpc_send and mpc_recv yields a round-trip latency of 88 s.