3.14 Merge window part 2 [LWN.net]

3.14 Merge window part 2

Posted Jan 30, 2014 3:26 UTC (Thu) by dlang (guest, #313)
Parent article: 3.14 Merge window part 2

thanks for the note on TCP autocorking, this is a sysctl item that I will need to remember to set by default from now on.

If the system is idle, what does it hurt to send small packets. If it's busy and small packets start to clog things up, the packets will end up getting consolidated anyway.

to post comments

3.14 Merge window part 2

Posted Jan 30, 2014 4:13 UTC (Thu) by Aliasundercover (guest, #69009) [Link] (3 responses)

I frequently turn the Nagle algorithm off with setsockopt(... TCP_NODELAY ...) when dealing time sensitive communications between machines in automation. I wonder if TCP autocorking will mean I need to tweak another knob. I certainly hope not.

3.14 Merge window part 2

Posted Jan 30, 2014 9:16 UTC (Thu) by cwillu (guest, #67268) [Link] (2 responses)

> tcp_autocorking - BOOLEAN
> Enable TCP auto corking :
> When applications do consecutive small write()/sendmsg() system calls,
> we try to coalesce these small writes as much as possible, to lower
> total amount of sent packets. This is done if at least one prior
> packet for the flow is waiting in Qdisc queues or device transmit
> queue. Applications can still use TCP_CORK for optimal behavior
> when they know how/when to uncork their sockets.
> Default : 1

I read this as saying that corking will only happen if you already had packets waiting in the outgoing buffer anyway.

> * Because TX completion will happen shortly, it gives a chance
> * to coalesce future sendmsg() payload into this skb, without
> * need for a timer, and with no latency trade off.

3.14 Merge window part 2

Posted Jan 30, 2014 10:25 UTC (Thu) by dlang (guest, #313) [Link] (1 responses)

if there are already packets waiting in the outgoing buffer, won't that delay the sending of the packets so they can be consolidated anyway? so why would auto-corking be needed?

3.14 Merge window part 2

Posted Feb 13, 2014 16:30 UTC (Thu) by Jan_Zerebecki (guest, #70319) [Link]

Handing off a new packet for transmission (uncork) can be delayed without delaying transmission of another packet that was already handed off.