fixes #1663 Request/Reply Protocol Throughput and Scalability

This eliminates the req protocols use of nni_timer (and setting
a single timer node per request.  This was problematic because it
devolves into O(n^2) as we wind up inserting timer nodes and having
to scan the list for the timer node.

The solution is to use a single scan - stop worrying about insertion,
but instead use a coarse granularity timer (defaults to 1 second)
for retries.  Then do the O(n) scan just once per interval.

A new option, NNG_OPT_REQ_RESENDTICK, can be used to change the tick
interval for cases (like unit tests) where more fine grained timing
is required.

1 files changed, 7 insertions, 0 deletions

diff --git a/src/compat/nanomsg/nn.c b/src/compat/nanomsg/nn.c
index 0f975b2d..af4db5d3 100644
--- a/src/compat/nanomsg/nn.c
+++ b/src/compat/nanomsg/nn.c
@@ -1144,6 +1144,13 @@ nn_setsockopt(int s, int nnlevel, int nnopt, const void *valp, size_t sz)
 		nn_seterror(rv);
 		return (-1);
 	}
+	if ((nnlevel == NN_REQ) && (nnopt == NN_REQ_RESEND_IVL)) {
+		// Only one context here, so it won't be too bad to tick
+		// as quickly as this, and it avoids some possible friction
+		// (e.g. with legacy tests).
+		(void) nng_socket_set_ms(sid, NNG_OPT_REQ_RESENDTICK, 10);
+	}
+
 	return (0);
 }