git: kernel - Fix issue with ARP packets stalling out entire network

Matthew Dillon dillon at crater.dragonflybsd.org
Thu Feb 21 15:39:43 PST 2013


commit b069f9cd471adb5001fcf32854e273cb92448641
Author: Matthew Dillon <dillon at apollo.backplane.com>
Date:   Thu Feb 21 15:33:26 2013 -0800

    kernel - Fix issue with ARP packets stalling out entire network
    
    * ARP packets can cause ARP routing table updates to occur.  An ARP
      routing table update is an expensive synchronous netmsg that is
      forwarded through *ALL* cpus.
    
    * ARP was previously being handled by netisr 0 and on large multi-way
      machines (aka monster the 48-way opteron) under very heavy loads this
      could result in very long stalls for any packet processing forwarded
      to cpu 0.
    
      Stalls exceeding 200 seconds were observed on monster when a large
      number of ARP packets had to be processed.
    
    * Implement a dedicated thread feature for the NETISR mechanism and
      modify NETISR_ARP to use it.  This takes the expensive synchronous
      ARP packet processing off the general per-cpu netisr threads.
      This thread currently runs on cpu (18 % ncpus) (NETISR_ARP == 18).
    
      Thus the general per-cpu (netisr 0) thread will no longer stall
      on ARP packets.
    
    * ping latencies under extreme loads improved to (approximately):
    
      ping -i 0.001 monster-nr
      11735 packets transmitted, 11735 packets received, 0.0% packet loss
      round-trip min/avg/max/stddev = 0.073/0.190/27.019/0.382 ms

Summary of changes:
 sys/net/netisr.c       | 30 ++++++++++++++++++++++++++++--
 sys/net/netisr.h       |  2 ++
 sys/netinet/if_ether.c | 17 ++++++++++++++++-
 3 files changed, 46 insertions(+), 3 deletions(-)

http://gitweb.dragonflybsd.org/dragonfly.git/commitdiff/b069f9cd471adb5001fcf32854e273cb92448641


-- 
DragonFly BSD source repository



More information about the Commits mailing list