Mailing List Archive

Hi,

Alexander> Hi...

Alexander> i got a Backhand cluster running for www.chefkoch.de
Alexander> http://www.chefkoch.de/backhand/

Alexander> The cluster is build as proxy cluster so that all questions
Alexander> got to the cook1 server. sometimes the cook1 serves some
Alexander> of the requests. Is it possible to take out the
Alexander> cook1 of serving request so that he does only the proxying ?

Alexander> here is the config of cook 1.

Alexander> <Directory /www/httpd/chefkoch.de/html/>
Alexander> # ... stuff ...
Alexander> Backhand byAge 3
Alexander> Backhand byRandom
Alexander> Backhand byLoad
Alexander> </Directory>


Alexander> UnixSocketDir /usr/local/apache/backhand
Alexander> MulticastStats 213.70.6.16 213.70.6.255:4445
Alexander> BackhandConnectionPools Off
Alexander> AcceptStats 213.70.6.0/24

Alexander> is there a way to let the broardcast only got to the cluster servers not
Alexander> to the intire class-c net ?

Using a multicast address instead of the broadcast address should do
that trick. The message will be sent on the network but only the
cluster servers (that accept messages on this multicast address)
network interface will pass the message to the operating system on
these cluster servers. For the rest of the class-c computers, the
message will be masked at the interface level (before getting to the
operating system). 225.213.70.6 is an example of a valid multicast address
and should be used instead of the 213.70.6.255 in your example.

Alexander> Is there a way to get more performance out of the cluster ?

Alexander> Thanks for helping.

Alexander> Regards

Alexander> Alex





Alexander> _______________________________________________
Alexander> backhand-users mailing list
Alexander> backhand-users@lists.backhand.org
Alexander> http://lists.backhand.org/mailman/listinfo/backhand-users

Hi...

> Using a multicast address instead of the broadcast address should do
> that trick. The message will be sent on the network but only the
> cluster servers (that accept messages on this multicast address)
> network interface will pass the message to the operating system on
> these cluster servers. For the rest of the class-c computers, the
> message will be masked at the interface level (before getting to the
> operating system). 225.213.70.6 is an example of a valid multicast
address
> and should be used instead of the 213.70.6.255 in your example.

Is Multicasting working on a "normal" cisco switched Network or does i have
to implement multicasting in the hardware/software ?

Regards

Alex

Alexander Meis wrote:

>Is Multicasting working on a "normal" cisco switched Network or does i have
>to implement multicasting in the hardware/software ?
>
>
Multicast should "work". However, to make it better than broadcast,
your switch needs to be multicast-aware and support "IGMP snooping".
2924 Cats do not support this, but I think every Cat that runs CatOS does.

--
Theo Schlossnagle
1024D/82844984/95FD 30F1 489E 4613 F22E 491A 7E88 364C 8284 4984
2047R/33131B65/71 F7 95 64 49 76 5D BA 3D 90 B9 9F BE 27 24 E7

> Multicast should "work". However, to make it better than broadcast,
> your switch needs to be multicast-aware and support "IGMP snooping".
> 2924 Cats do not support this, but I think every Cat that runs CatOS
does.

hmmm so i need to change the switch for doing the multicast ?
or can i upgrade the cisco os on the 2924 ?
Even it is a 2924-EN

Greetings

alex

Theo,

I don't think this is an accurate representation.
Even if the Cisco switch will treat this as broadcast,
as far as the servers are concerned, using multicast will be much better
than broadcast for the reason explained in my earlier message.

Cheers,

:) Yair.

Theo> Alexander Meis wrote:

>>Is Multicasting working on a "normal" cisco switched Network or does i have
>>to implement multicasting in the hardware/software ?
>>
>>
Theo> Multicast should "work". However, to make it better than broadcast,
Theo> your switch needs to be multicast-aware and support "IGMP snooping".
Theo> 2924 Cats do not support this, but I think every Cat that runs CatOS does.

Yair Amir wrote:

>Theo,
>
>I don't think this is an accurate representation.
>Even if the Cisco switch will treat this as broadcast,
>as far as the servers are concerned, using multicast will be much better
>than broadcast for the reason explained in my earlier message.
>
>
I agree with you that it is better for the servers. However, you still
send packets to all machines on the attached switch if the switch isn't
capable of IGMP snooping. I thought the original concern was with
network bandwidth.

Assuming I have a 100Mbs switch;
If I have machines A B and C in a Spread ring pushing 40 Mbs. and I have
machines D and E trying to send 80 Mbs over TCP/IP to each other, you
can see the problem. If the switch isn't smart enough to send the 40Mbs
of multicast traffic over _only_ the ports that A B and C are attached
to, then I can't send 80 Mbs between D and E. Good switches perform
IGMP snooping and will only propagate the multicast traffic to concerned
parties which will free up the bandwidth on individual ports of
unconcerned parties.

This is a serious issue in any environment with regular backup
schedules. The backup machine wants to pull data from client hosts at
full bandwidth (100Mbs). If ever host is being sent multicast IP
traffic (including the backup host) you are filling the pipe with lots
of useless packets that just get in the way.

And to answer someone else's question... I don't think a 2924XL-EN will
do IGMP snooping, but feel free to correct me. An Avaya Cajun P334T
will do it and it is about the same price.

--
Theo Schlossnagle
1024D/82844984/95FD 30F1 489E 4613 F22E 491A 7E88 364C 8284 4984
2047R/33131B65/71 F7 95 64 49 76 5D BA 3D 90 B9 9F BE 27 24 E7