Update collectd and graphite UDP listeners with perf enhancements

[sparrc]

closes #4678

related to #4676

cc @otoolep

[sparrc]

changed this just for consistency with graphite and udp services

[sparrc]

cc @dswarbrick @daviesalex @nkatsaros

[otoolep]

We need the CHANGELOG updated, and https://github.com/influxdb/influxdb/blob/master/etc/config.sample.toml updated to include these new values.

[otoolep]

CHANGELOG should be updated in the "Features" section -- allow read buffer size to be set.

[otoolep]

This is a little vague. Why not be precise and repeat the text in the Go docs?

"Sets the size of the operating system's receive buffer associated with the UDP traffic "

[sparrc]

kk 👍

[otoolep]

Does the buffer size apply to TCP traffic? If so, we should set it for Graphite TCP, right?

[sparrc]

For graphite we can't change the tcp read buffer at the moment. We would have the change the net.Listen call to a net.DialTCP call to get that functionality.

[dswarbrick]

Somewhere deep down, SetReadBuffer() will be doing the equivalent of a setsockopt() with SO_RCVBUF option. Note that Linux (presumably other *nixes) has maximums that this can be set to. The following are from my workstation, running kernel 4.2:

net.core.rmem_default = 212992
net.core.rmem_max = 212992
net.ipv4.tcp_rmem = 4096    87380   6291456

These can of course be modified with sysctl -w ...

[nkatsaros]

You can do tcpConn, ok := conn.(*net.TCPConn) and do whatever you want on the connection that s.ln.Accept() spits out. There's no need to stray from the generic Listen/Accept/Dial functions in the net package.

[sparrc]

@dswarbrick That is also what I get:

$ sudo sysctl -a | grep net.core.rmem
net.core.rmem_default = 212992
net.core.rmem_max = 212992

maybe it can adjust the max? I'm not sure...

[sparrc]

@nkatsaros, from what I understand, net.Listen("tcp"...) is a simplified alternative to running net.DialTCP(...). I was just saying that it would be easier to adjust this if we used DialTCP, I suppose we could typecast and adjust on every for-loop iteration but that seems messy.

[dswarbrick]

@sparrc Need to be root to modify the OS defaults / max. These should not be fiddled with directly by InfluxDB, as they will affect all future sockets, in all applications. It would be worth documenting how to change them, or linking to relevant documentation.

[sparrc]

oh, looks like BSD already has the max set higher, that would explain why it worked

OSX 10.11 (8MB)

$ sysctl kern.ipc.maxsockbuf
kern.ipc.maxsockbuf: 8388608

FreeBSD 10.1 (2MB)

$ sysctl kern.ipc.maxsockbuf
kern.ipc.maxsockbuf: 2097152

Get it together Linux!!

I'll put a note in the documentation

[otoolep]

Thanks for the heads-up @dswarbrick

@sparrc -- looks like we have documentation issue here.

[sparrc]

@otoolep finally tracked down the BSD/Darwin failures I was seeing related to this. This PR should be good to merge now.