tcp: Aggregate the mbuf in sosendtcp() a little bit

This greatly reducse the ipi interrupts caused by ipi sending (both
domsg and sendmsg), thus improves overall performance a bit.

net.inet.tcp.sosnd_agglim is added to tune how much mbuf should be
aggregated; it is default to 2.  Setting this sysctl to 1 restores
the old behaviour: one full mbuf at a time.

On Phenom 9550 (4 core, 2.2GHz):
8 parallel netperf -H 127.0.0.1 -P0 (4 runs, unit; Mbps)

                                              IPIs/s (sum of 4 core)
 1 mbuf  6735.98  6903.13  6971.89  7056.66   ~400K
 2 mbuf  7675.47  7757.28  7815.45  7514.50   ~240K
 4 mbuf  7895.33  7584.22  7704.12  7723.33   ~180K
 8 mbuf  8006.94  8077.87  7701.23  8061.12   ~120K
16 mbuf  8151.68  8023.03  7972.42  8046.13   ~100K

The default value (2) for the sosnd_agglim improve the whole
performance by ~10%.  IPI rate is also reduce greatly.

It has no obvious impact on 1000BaseT or 100baseTX network performace.

diff --git a/sys/kern/uipc_socket.c b/sys/kern/uipc_socket.c
index 924839d..f8eb2c7 100644 (file)
--- a/sys/kern/uipc_socket.c
+++ b/sys/kern/uipc_socket.c
@@ -99,6 +99,8 @@
 
 #include <machine/limits.h>
 
+extern int tcp_sosnd_agglim;
+
 #ifdef INET
 static int      do_setopt_accept_filter(struct socket *so, struct sockopt *sopt);
 #endif /* INET */
@@ -910,6 +912,8 @@ restart:
                }
                mp = &top;
                do {
+                   int cnt = 0;
+
                    if (uio == NULL) {
                        /*
                         * Data is prepackaged in "top".
@@ -936,7 +940,8 @@ restart:
                        mp = &m->m_next;
                        if (resid == 0)
                                break;
-                   } while (space > 0 && 0 /* XXX */);
+                       ++cnt;
+                   } while (space > 0 && cnt < tcp_sosnd_agglim);
 
                    if (flags & MSG_OOB) {
                            pru_flags = PRUS_OOB;

diff --git a/sys/netinet/tcp_input.c b/sys/netinet/tcp_input.c
index 4535dfc..7bb4d5f 100644 (file)
--- a/sys/netinet/tcp_input.c
+++ b/sys/netinet/tcp_input.c
@@ -227,6 +227,9 @@ int tcp_autorcvbuf_max = 2*1024*1024;
 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_RW,
     &tcp_autorcvbuf_max, 0, "Max size of automatic receive buffer");
 
+int tcp_sosnd_agglim = 2;
+SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosnd_agglim, CTLFLAG_RW,
+    &tcp_sosnd_agglim, 0, "TCP sosend mbuf aggregation limit");
 
 static void     tcp_dooptions(struct tcpopt *, u_char *, int, boolean_t);
 static void     tcp_pulloutofband(struct socket *,