tcp mss MCLBYTES restriction

[Andrew Gallatin]

Andre Oppermann writes:
 > 
 > Could you run some bechmarks with the current MCLBYTES rounding
 > and without it on 100Mbit 1.5kMTU and GigE with 9k MTU?

David Borman is totally right.   Clipping the mss is really worth it,
especially with zero-copy sockets.   Forget I said anything.

Here is some data I just took..  Send and receiver are 3GHz P4s.
Test is netperf -- -S131072 -s131072.  Sender is running current
as of Saturday.  Receiver is running 4.9-Release.

"em" is Intel(R) PRO/1000 Network Connection, Version - 1.7.19
"myri" is a M3F2-PCIXE-2. (4Gb/s).

NIC    mtu     BW     %cpu	mss_truncate	zero_copy
       (bytes) (Mb/s) 
----   ------  -----  ----	-----		------
em     16114   727    12.8	0		0
em     16114   730    15.2	1		0
em     16114   728    12.4	0		1
em     16114   743     7.3	1		1
em     9000    870    19.2	0		0
em     9000    900    21.6	1		0
em     9000    867    18.2	0		1
em     9000    845     8.9	1		1
myri   9000    1972   69.0	0		0
myri   9000    2312   68.6	1		0
myri   9000    2609   56.3	0		1
myri   9000    3009   52.3	1		1



Drew