2

As I was running apachebench on a nginx -> node.js server, I noticed after a certain number of concurrent requests were set, some of the requests were not being followed through to completion. The results of ss -s output are as follows:

TCP:   26006 (estab 660, closed 25296, orphaned 38, synrecv 0, timewait 25295/0), ports 0

It looks like the majority of the connections have been closed and are waiting. I imagine this will work well under a heavy load.

Is it possible to "recycle" these connections faster? They seem to persist for quite a long time (10 seconds or so?)

dsp_099
  • 5,801
  • 17
  • 72
  • 128

2 Answers2

0

They persist for two minutes actually. They are not 'requests' but connections, and they have all closed.

You should make sure you have a reasonable HTTP keepalive interval set so that peers can reuse TCP connectors for multiple HTTP requests.

The TIME-WAIT state is an important part of TCP security and shouldn't be messed with.

user207421
  • 305,947
  • 44
  • 307
  • 483
0

This may help explain why TIME_WAIT exists and how these connections end up in TIME_WAIT in the first place.

I'm interested to know if you're running the test on a single machine, if so, it could be either the client side or the server side that is leaving the sockets in TIME_WAIT. If it's the client side of the test then you don't really need to worry that much. If you are running the test on one machine split it so you have the server on its own so that you can see what's generating the TIME_WAITs.

If the TIME_WAITS are being generated by the server then you need to work out why you are initiating the active close on the server and see if it's practical to move that active close to the client.

Len Holgate
  • 21,282
  • 4
  • 45
  • 92