[ https://issues.redhat.com/browse/JGRP-2504?page=com.atlassian.jira.plugin... ] Andrew Skalski commented on JGRP-2504: -------------------------------------- This change doesn't appear to be working yet.  I looked at what it's doing using strace: {code:java} $ strace -f -o strace.out java -Djgroups.tcpping.initial_hosts=jgroups-west[7800],jgroups-east[7800] -cp jgroups-5.1.0.Alpha1-SNAPSHOT.jar:. SpeedTest <<< quit{code} The receive buffer is still not being configured ahead of bind/listen: {code:java} 7241 socket(AF_INET6, SOCK_STREAM, IPPROTO_IP) = 9 7241 setsockopt(9, SOL_IPV6, IPV6_V6ONLY, [0], 4) = 0 7241 fcntl(9, F_GETFL) = 0x2 (flags O_RDWR) 7241 fcntl(9, F_SETFL, O_RDWR|O_NONBLOCK) = 0 7241 setsockopt(9, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0 7241 pread64(4, "\312\376\272\276\0\0\0007\0W\n\0002\0003\t\0\4\0004\v\0005\0006\7\0007\n\08\0"..., 1918, 9994511) = 1918 7241 pread64(4, "\312\376\272\276\0\0\0007\0/\n\0\10\0$\t\0\7\0%\n\0\t\0&\n\0\t\0'\7\0"..., 985, 16509997) = 985 7241 pread64(4, "\312\376\272\276\0\0\0007\1]\n\0Y\0\272\n\0\273\0\274\10\0\275\n\0}\0\276\t\0\30\0"..., 7728, 16686562) = 7728 7241 pread64(4, "\312\376\272\276\0\0\0007\0\33\n\0\3\0\26\7\0\27\7\0\30\1\0\6<init>\1\0"..., 536, 16509461) = 536 7241 bind(9, {sa_family=AF_INET6, sin6_port=htons(7800), inet_pton(AF_INET6, "::ffff:69.164.216.53", &sin6_addr), sin6_flowinfo=htonl(0), sin6_scope_id=0}, 28) = 0 7241 listen(9, 50) = 0 {code} Looking at the code, note [this section|https://github.com/belaban/JGroups/blob/master/src/org/jgroups/pr...]; the TcpServer constructor (which tries to pass recv_buf_size along to Util.createServerSocket) is run before recv_buf_size is set. Addressing the question of why the buffer size needs to be configured on the listening socket (rather than the accepted socket): Although I haven't yet read through the kernel sources to fully understand what's going on, I believe at least part of it has to do with negotiating the window scaling TCP option, which happens during connection handshake. By the way, if you don't have two geographically separated servers, you can create artificial latency using the Linux Netem module.  (The kernel module name is sch_netem and is included in the kernel-modules-extra RPM; you will also need the iproute-tc RPM.)  The syntax is arcane, so I will attach a shell script to this ticket. One other thing: I'm not sure where's the best place to report this, but the "Bug Reports" page on jgroups.org links to an old Jira URL at jira.jboss.com that no longer works. -- This message was sent by Atlassian Jira (v7.13.8#713008)