[Spread-users] Spread segments w/o active daemons cause delays in other segments

[Ryan Caudy]

Under most circumstances, the timeouts controlled by Wide_network only
affect the daemon membership algorithm.  Additionally, Hurry_timeout
has some affect on a network that hasn't been active.

In your tests, have the daemons completed the installation of their
initial membership configuration before you start the flooders?  This
would have a major affect, but would only reflect a less common case
of Spread's behavior.

Cheers,
Ryan

On Fri, 18 Mar 2005 12:07:31 -0500, Dmitry Korsun <dkorsu at corp.idt.net> wrote:
> Ryan,
> 
>     Thanks for the quick reply.  I profiled the suggested configuration
> and found no performance drop in local spread message transfer rate (the
> 3rd machine does not run a spread daemon).  After a few tests I think I
> traced the delay cause down to the timeouts defined in membership.c -
> they are adjusted based on Wide_network flag.  From what I understand
> the Timeouts are increased if more than 1 segment is defined and at
> least one segment services a different subnet.  This is exactly the
> scenario found in my configuration.
>     Just to confirm, I recompiled spread and hardcoded Wide_network=0.
> The delay phenomenon described below disappeared.  Is there a way to
> determine which timeout is responsible for such behavior (other than
> adjusting each one at a time and re-running the tests) ?
> 
>     On a separate topic, I came across "Flow Control" settings for
> spread daemons.  However, the spread documentation barely mentions it
> and does not go into detail about the values - window size, and the
> values assigned to each of the daemons (that's how spmonitor configures
> the flow control parameters).
> 
> Any help would be greately appreciated,
> Thanks.
> 
> -Dmitry K.
> 
> Ryan Caudy wrote:
> 
> >I can't think of anything that would cause this performance drop.
> >Just out of curiosity, can you try adding a third daemon to the first
> >segment in the configuration file, removing the second segment, and
> >repeating the tests with the same two daemons up and participating?
> >
> >Cheers,
> >Ryan
> >
> >
> >On Thu, 17 Mar 2005 23:37:39 -0500, Dmitry Korsun <dkorsu at corp.idt.net> wrote:
> >
> >
> >>Hi all.
> >> I am relatively new to Spread and have stumbled across a problem that I
> >>need help with resolving. Apparently adding a second Spread segment
> >>without any spread daemons causes a big performance drop in the first
> >>segment, even when clients are connecting to a member of the first
> >>segment locally.
> >>
> >>Here are the scenarios:
> >>
> >>I was getting really good performance results when using a single
> >>segment configuration with spread daemons runing on BOTH machines in the
> >>segment:
> >>#
> >>Spread_Segment 172.16.202.255:4803 {
> >>       DevLinuxPro     172.16.202.64
> >>       StarDev1 172.16.202.77
> >>}
> >>#
> >>Performance was tested by using spflooder utility and connecting to the
> >>spread daemon LOCALLY:
> >>devlinuxpro$ time spflooder -u ya -m 10000 -b 100 -s 4803 at devlinuxpro
> >>flooder: connecting to 4803 at devlinuxpro.mis.idt.net
> >>flooder: starting  multicast of 10000 messages, 100 bytes each.
> >>flooder: completed  10000 messages of 100 bytes
> >>flooder: completed multicast of 10000 messages, 100 bytes each.
> >>real    0m0.559s
> >>user    0m0.020s
> >>sys     0m0.240s
> >>
> >>This yields ~20000 messages/sec
> >>
> >>Once a second segment was configured but no daemon started on the new
> >>segment machine, performance dropped dramatically:
> >>#
> >>Spread_Segment 169.132.9.255:4803 {
> >>       dev_1       169.132.9.113
> >>}
> >>#
> >>Again, performance was tested LOCALLY on the machine that is part of the
> >>first segment:
> >>devlinuxpro$ time spflooder -u ya -m 10000 -b 100 -s
> >>4803 at devlinuxpro.mis.idt.net
> >>flooder: connecting to 4803 at devlinuxpro.mis.idt.net
> >>flooder: starting  multicast of 10000 messages, 100 bytes each.
> >>flooder: completed multicast of 10000 messages, 100 bytes each.
> >>real    0m25.543s
> >>user    0m0.060s
> >>sys     0m0.960s
> >>Now we are down to ~400 messages/sec.
> >>
> >>I about to test performance once a spread daemon is brought up on the
> >>second segment.
> >>
> >>Question - why does this addition of a segment without a daemon process
> >>affect performance on a entirely different segment, especially the LOCAL
> >>performance.  Is there a way to configure SPREAD so that different
> >>segments do not affect each other's performance ?
> >>
> >>Thank you in advance,
> >>Dmitry Korsun
> >>
> >>--
> >>::::::::::::...... .  .   .
> >>:: Dmitry Korsun
> >>:: x4095
> >>:: dmitry.korsun at corp.idt.net
> >>:::::::::::::::::....... .   .       .
> >>
> >>--
> >>::::::::::::...... .  .   .
> >>:: Dmitry Korsun
> >>:: x4095
> >>:: dmitry.korsun at corp.idt.net
> >>:::::::::::::::::....... .   .       .
> >>
> >>_______________________________________________
> >>Spread-users mailing list
> >>Spread-users at lists.spread.org
> >>http://lists.spread.org/mailman/listinfo/spread-users
> >>
> >>
> >>
> 
> --
> ::::::::::::...... .  .   .
> :: Dmitry Korsun
> :: x4095
> :: dmitry.korsun at corp.idt.net
> :::::::::::::::::....... .   .       .
> 
>