[Spread-users] Sudden freezes in message delivery using C API

[John Schultz]

On Wed, 23 May 2007, Doug Palmer wrote:
> I've timed the freezes and it works out to be 2 seconds every time. So
> token loss looks like a likely candidate.

The quickest diagnosis then would be to simply lower the Hurry_timeout and 
see if that affects your hiccup timings.

> We have a peculiar setup where all the machines appear to be on a single
> virtual LAN but we observed this behaviour before we separated the two
> ends. So I'd like do a test where I put each machine in its own segment
> and see whether . Is is possible to set up a spread.conf where machines
> on the same subnet are in separate segments?

Yes, simply put each machine in its own segment, the broad/multicast 
addresses don't matter.  I'm not sure what this will tell you though -- 
Spread decides on whether you are a wide network or not by looking at all 
of the machines addresses -- if they are all in the same class B \16 
network, then it assumes a LAN, otherwise a WAN.

A potential fix to your problem is to send the token multiple times on the 
network when you logically forward it to the next member.  However, if an 
intermediary router is dropping your packets because its buffers are full 
from your burst, then this might not solve your problem either.

Another potentiality is that you have a bad network hardware that is 
causing lots of loss on a particular link.  You can look at the output of 
the spmonitor prgram when it is reporting status on all the machines and 
if you see lots of retrans, then you might have a network problem that 
Spread is overcoming with some work.

---
John Schultz
Spread Concepts
Phn: 443 838 2200