Hi everybody,

  I am experiencing RX packet drops sometimes and it seems that TIPC has 
some hard time dealing with that. When I looked in tipc_link.c I found 
the following comment:

/* TODO: Implement stronger sequence # checking someday ... */

and .. then I got a bit worried :)
What exactly is the expected behavior in case of droped eth packets?
As soon as the nic drops even few packets, the TIPC stack becomes .. 
well completely unpredictable. In 80% of the cases a disable/enable of 
the bearer helps, but there are those 20% when only complete reboot 
solves the problem. Also affected links never timeout/reset on their 
own. I am working on resolving the packet drop issue in general, but 
that's not a long term solution, neither is trying to detect if things 
went wrong. It's also worth noting that those 20% of cases tend to 
happen when my apps try to transfer 2-3MB or more in series of 16K 
send()s over a SOCK_SEQPACKET socket.

About the setup:
* TIPC network of 20+ nodes, Intel/AMD, 32bit, SMP all of them
* NICs are either e1000 or tg3 driven, working over a gbit full duplex 
link, all nodes are connected to a 24 port gbit d-link switch
* TIPC is 1.7.5 with two of the 4 posted/available patches:
  - Prevent-premature-discarding-of-messages-during-fragment-reassembly
  - Fix-port-related-bugs-arising-when-TIPC-network-address-is-assigned
* issue exists with both 2.6.20.4 and 2.6.24.2 kernels
* tested with NIC qlen of 1000 - 8192, tipc link window up to 4096 packets
* under normal conditions max send queue stays within the 60-100 range.

One more note: I had similar problems with the stock TIPC 1.6.2 some 
time ago. Back then the stack was much more predictable and a 
disable/enable of the bearer helped in 99% of the cases. I resolved the 
issue by increasing tipc link window (a lot .. to 1024 at first) and 
upgrading to TIPC 1.7.5 (which btw manages 2+ times smaller max send 
queue under the same conditions .. thanks for that ;) .. 1.7.5 seems to 
work much more smoothly in general).

Any help/info/suggestions?

Regards,
Peter Litov.


-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2008.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/
_______________________________________________
tipc-discussion mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/tipc-discussion

Reply via email to