Hello
The round robin interface usage applies only to outgoing packets, incoming
packets is still a matter of the MAC address of the interface and the resolution
of IP address to MAC address by the sender or the local gateway's ARP tables.
By default all Sun network Interfaces have the same MAC address (this can be
changed in the eeprom) which means that all interfaces receive the same packets
into their buffers if they are connected to the same physical segment so there
is very little performance gain in receiving network traffic. If you do set
use local MAC addresses in the eeprom, you then have the issue of how to get the
originating machine or the gateway to load share across the interfaces (my
experience with this setting has been mixed, some Sun's observe it, some do not;
so I usually end up mangling the MAC addresses myself with ifconfig). Once you
have each interface on a separate MAC address, depending on Solaris's
roundrobin over routes will work if you target an IP address that would have the
same hop count over all the interfaces that you are sharing over, otherwise it
would just go to the interface with the IP address targeted, ie.: you need to
target an IP address configured on the target host but not on the subnet over
which you are trying to communicate to it with. For example, if you are trying
to share bandwidth with routes over the four interfaces of a quad interface
card, you would direct traffic to the IP address of a fifth interface (probably
the motherboard one) and so long as that fifth interface is on another subnet
than the four of the quad card, the routes would be the same metric over all
four interfaces (one way would be to plumb a virtual interface on a subnet that
you never use).
Regards,
John I Wang
Sr. Systems Engineer
Steverson Information Professionals
---
Enron Broadband Services
3 Allen Center, Room 337C
PH (713) 345-6238
|--------+----------------------->
| | BLumpkin@echa|
| | rge.com |
| | |
| | 10/16/00 |
| | 03:10 PM |
| | |
|--------+----------------------->
>----------------------------------------------------------------------------|
| |
| To: jfielden AT excitecorp DOT com, steingra AT pprd.abbott DOT
com |
| cc: ravi.channavajhala AT csfb DOT com,
|
| veritas-bu AT mailman.eng.auburn DOT edu, (bcc: John
Wang/Contractor/Enron |
| Communications) |
| Subject: RE: [Veritas-bu] Network load balancing |
>----------------------------------------------------------------------------|
Solaris 2.6 will actually round-robin its interface usage if more than one
is found to have the same metric and that metric is the lowest route to the
destination. I remember having to disable this functionality because it was
screwing up the firewall when I was at The Seattle Times.
The Kernel Parameters are something like enable_group_ifs and I beleive
enable_strict_dst_multihoming.
something like that.
--Buddy
-----Original Message-----
From: Joshua Fielden [mailto:jfielden AT excitecorp DOT com]
Sent: Monday, October 16, 2000 11:50 AM
To: Andrew Steingruebl
Cc: Ravi Channavajhala; veritas-bu AT mailman.eng.auburn DOT edu
Subject: Re: [Veritas-bu] Network load balancing
On Mon, Oct 16, 2000 at 01:36:16PM -0500, Andrew Steingruebl filled up my
inbox with:
>
> Ravi Channavajhala said:
>
> >Is it possible to do network load balancing across several
> >physical interfaces in a media server i.e., one storage
> >unit being shared by several netbackup slave server
> >definitions eg., backup-e0, backup-e1, backup-e2? The
> >primary definition is backup (nodename). All these interfaces
> >are on different subnets. The storage unit is defined for
> >backup instance only.
> >
> >I already defined these in bp.conf on the master as well as
> >the slave. What I dont understand is, how does the master
> >make a determination to direct a specific backup stream to
> >a specific network interface. Thanks.
> >
>
> There are 2 streams of traffic for each client.
> - backup stream
> - catalog stream
>
> A server will automatically talk to another machine using the interface
with
> the best metric. On Unix, if you've got 2 interfaces, the system will use
the
> interface on the network its trying to reach.
>
If only this were true for every platform. On Solaris, this is not the truth
in practice. Also, if you have an environment like we have, with 11
interfaces trying to take 'least hop' to 30 changing VLAN's, you know that
static routing is not a scalable solution, nor is 30 interfaces in a
machine.
> The client will need to know about each server name. So, what you
probably
> want to do is:
>
> 1. Set up multiple "A" records/IP's for your master server.
> 2. Set up 1 storage unit and makes it owned by whatever interface name you
> consider primary.
> 3. List all of the "aliases" for the master server in the bp.conf file on
> master and slaves.
>
> All backup traffic will take the optimal route. All catalog traffic must
get
> sent to "master" which is the top line in the bp.conf file.
>
> You will have asymmetric traffic flows, but your backup data will tape the
> optimal route.
>
And if data goes up one interface, and down another for any reason (router
failure, hsrp kicks in, different path is taken...), the software recovers
less than gracefully. 'Chokes' is a better term.
With that said, it Does The Right thing(tm) most of the time, just not all
of the time, and not as simply as stated.
--
"Any man page that includes the words "USE AT YOUR OWN RISK. BEWARE OF DOG.
SLIPPERY WHEN WET" means trouble" - Michael Lucas
Joshua Fielden, Senior Systems Administrator and Backups Team Lead
eXcite@Home, Inc. jfielden AT excitecorp DOT com 650-556-3316
_______________________________________________
Veritas-bu maillist - Veritas-bu AT mailman.eng.auburn DOT edu
http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
_______________________________________________
Veritas-bu maillist - Veritas-bu AT mailman.eng.auburn DOT edu
http://mailman.eng.auburn.edu/mailman/listinfo/veritas-bu
|
