File: README.Solaris

package info (click to toggle)
swish%2B%2B 6.1.5-2
  • links: PTS
  • area: main
  • in suites: squeeze
  • size: 2,256 kB
  • ctags: 1,759
  • sloc: ansic: 11,931; lisp: 804; sh: 629; perl: 366; makefile: 80
file content (57 lines) | stat: -rw-r--r-- 2,021 bytes parent folder | download | duplicates (9)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
	Intent
	======
	This note is about running search(1) as a daemon under Solaris
	where it will be expected to service a large number of requests
	per unit time.


	Background
	==========
	The first side of a TCP connection to close it goes into the
	TIME-WAIT state.  From RFC 1122:

		The graceful close algorithm of TCP requires
		that the connection state remain defined on (at
		least) one end of the connection, for a timeout
		period of 2xMSL, i.e., 4 minutes.  During this
		period, the (remote socket, local socket) pair
		that defines the connection is busy and cannot
		be reused.

	The request/response class of servers, such as web servers and
	search(1), tend to close the connection first.  From the above,
	this means that such a server's socket goes into the TIME-WAIT
	state.

	Under sustained heavy load, the server will run out of
	resources for sockets because they will be needed for new
	requests faster than they are coming out of TIME-WAIT.


	Solaris
	=======
	The Solaris TCP stack implements the quoted section of RFC 1122
	above to the letter.  The 4-minute wait can be considered too
	long by some.  To change that to, say, 1 minute, you can do:

		ndd -set /dev/tcp tcp_time_wait_interval 60000

	Solaris also starts ephemeral port allocation unusually high at
	32768.  Ephemeral ports can start at 1024.  To change this, you
	can do:

		ndd -set /dev/tcp tcp_smallest_anon_port 1024

	The above two changes will buy the server more time before it
	runs out of socket resources, but under sustained heavy load,
	it will still run out eventually.

	The best thing to do is to ensure that any search(1) clients
	you write call shutdown(2) after they have finished sending
	their request.  This will cause them to go into TIME-WAIT
	instead of the server.

	Note: this isn't a problem under Linux because, as far as I can
	tell, Linux violates RFC 1122 by allowing sockets in TIME-WAIT
	to be used for new connections.  Despite the violation, it
	works out for the best in the case of servers.