svn commit: r1539207 - /subversion/trunk/subversion/libsvn_ra_svn/marshal.c

[st...@apache.org]

Author: stefan2
Date: Wed Nov  6 01:11:31 2013
New Revision: 1539207

URL: http://svn.apache.org/r1539207
Log:
Revert r1533530.  It turns out that APR_TCP_NODELAY will have no effect
on real ethernet connections but only the loopback (localhost) device.
Tested with 'svn-bench null-list' over 1GbE and 10GbE on Linux.

Modified:
    subversion/trunk/subversion/libsvn_ra_svn/marshal.c

Modified: subversion/trunk/subversion/libsvn_ra_svn/marshal.c
URL: http://svn.apache.org/viewvc/subversion/trunk/subversion/libsvn_ra_svn/marshal.c?rev=1539207&r1=1539206&r2=1539207&view=diff
==============================================================================
--- subversion/trunk/subversion/libsvn_ra_svn/marshal.c (original)
+++ subversion/trunk/subversion/libsvn_ra_svn/marshal.c Wed Nov  6 01:11:31 2013
@@ -119,16 +119,6 @@ svn_ra_svn_conn_t *svn_ra_svn_create_con
             && apr_sockaddr_ip_get(&conn->remote_ip, sa) == APR_SUCCESS))
         conn->remote_ip = NULL;
       svn_ra_svn__stream_timeout(conn->stream, get_timeout(conn));
-
-      /* We are using large r/w buffers already.
-       * So, once we decide to actually send data, we want it to go over
-       * the wire a.s.a.p..  So disable Nagle's algorithm.
-       *
-       * We ignore the result of this call since it safe to continue even
-       * if we keep delaying.  The only negative effect is increased
-       * latency (can be additional 5 .. 10ms depending on circumstances).
-       */
-      apr_socket_opt_set(sock, APR_TCP_NODELAY, 1);
     }
   else
     {