A lot of testing and discussion regarding this issue occurred on irc. Here's a summary.
1. OSX reports very strange values when the 0mq library runs getsockopt for SO_SNDBUF. It returns 32. The thought is that it is returning kbytes instead of bytes like every other platform. :( 2. The assertion occurs because the system is running out of space in the send buffer. On OSX there are two separate send buffers; one is for TCP for remotes and the second is for local loopback communications. For symmetry, I have set both the send and recv buffers to larger values. # the sendspace + recvspace buffers must be less or equal to maxsockbuf kern.ipc.maxsockbuf=8388608 net.inet.tcp.sendspace=1048576 net.inet.tcp.recvspace=1048576 net.local.stream.sendspace=3000000 net.local.stream.recvspace=3000000 By default, OSX *ignores* these values unless a certain minimum number of sockets are open. To override that behavior and make sure that it *always* uses my values, do this: net.inet.tcp.sockthreshold=0 # default was 64 3. On my Mac, the maxsockbuf was stuck at 4MB. The only way to make it larger was to force the machine to boot with a 64-bit kernel (OSX defaults to a 32-bit kernel for most machines). The 64-bit kernel let me bump up the values a bit higher as can be seen above (maxsockbuf is set to 8MB). 4. OSX uses the launchctl daemon to override the 'max open files' (file descriptor) setting. Google for it. Note that you *cannot* pass "unlimited" as the hard limit; OSX *requires* that it be a number. After I made the buffer changes, everything is working okay. I am certain there is still a problem here particularly when the mailbox code needs to modify the send buffer size (SO_SNDBUF). That logic isn't working right. I'll make a ticket to track it. My test program can successfully create over 5k sockets now without bitching, so I'm happy. Thanks to mato, et al for help on irc. cr _______________________________________________ zeromq-dev mailing list zeromq-dev@lists.zeromq.org http://lists.zeromq.org/mailman/listinfo/zeromq-dev