git: tcp: Improve sender-sender and sender-receiver fairness on the same netisr

[Sepherosa Ziehau <sephe@xxxxxxxxxxxxxxxxxxxxxxx>]

commit 2fb3a851c77d952c5c8d64e64affd1c802adbc02
Author: Sepherosa Ziehau <sephe@dragonflybsd.org>
Date:   Thu Jan 17 16:33:55 2013 +0800

    tcp: Improve sender-sender and sender-receiver fairness on the same netisr
    
    Yield to other senders or receivers on the same netisr if the current TCP
    stream has sent certain amount of segments (currently 4) and is going to
    burst more segments.  sysctl net.inet.tcp.fairsend could be used to tune
    how many segements are allowed to burst.  For TSO capable devices, their
    TSO aggregate size limit could also affect the number of segments allowed
    to burst.  Set net.inet.tcp.fairsend to 0 will allow single TCP stream to
    burst as much as it wants (the old TCP sender's behaviour).
    
    "Fairsend" is performed at the places that do not affect segment sending
    during congestion control:
    - User requested output path
    - ACK input path
    
    Measured improvement in the following setup:
    
    +---+            +---+
    |   |<-----------| B |
    |   |            +---+
    | A |
    |   |            +---+
    |   |----------->| C |
    +---+            +---+
    
    A (i7-2600, w/ HT enabled), 82571EB
    B (e3-1230, w/ HT enabled), 82574L
    C (e3-1230, w/ HT enabled), 82574L
    The performance stats are gathered from 'systat -if 1'
    
    When A runs 8 TCP senders to C and 8 TCP receivers from B, sending
    performance are same ~975Mbps, however, the receiving performance before
    this commit stumbles between 670Mbps and 850Mbps; w/ "fairsend" receiving
    performance stays at 981Mbps.
    
    When A runs 16 TCP senders to C and 16 TCP receivers from B, sending
    performance are same ~975Mbps, however, the receiving performance before
    this commit goes from 960Mbps to 980Mbps; w/ "fairsend" receiving
    performance stays at 981Mbps stably.
    
    When there are more senders and receivers running on A, there is no
    noticable performance difference on either sending or receiving between
    non-"fairsend" and "fairsend", because senders are no longer being able
    to do continuous large burst.
    
    "Fairsend" also improves Jain's fairness index between various amount of
    senders (8 ~ 128) a little bit (sending only tests).

Summary of changes:
 sys/netinet/tcp_input.c  |  14 +++++--
 sys/netinet/tcp_output.c | 103 +++++++++++++++++++++++++++++++++++++++++++++--
 sys/netinet/tcp_sack.c   |  14 +++++++
 sys/netinet/tcp_subr.c   |   6 +++
 sys/netinet/tcp_usrreq.c |   4 +-
 sys/netinet/tcp_var.h    |  14 ++++++-
 6 files changed, 143 insertions(+), 12 deletions(-)

http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/2fb3a851c77d952c5c8d64e64affd1c802adbc02


-- 
DragonFly BSD source repository