Re: [load balancing] load balancing large farms?

From: Jeremy Noetzelman (jjnIZZATpeoplelink.com)
Date: Wed Oct 18 2000 - 15:21:38 EDT


Karyn Ulriksen wrote:
>
> Jeremy,
>
> Out of curiousity.... How many slots do you have populated on your CS800's
> and how many rulesets?
>
> Karyn
>
> > -----Original Message-----
> > From: Jeremy Noetzelman [mailto:jjnIZZATpeoplelink.com]
> > Sent: Tuesday, October 17, 2000 8:52 PM
> > To: lb-lIZZATvegan.net
> > Subject: Re: [load balancing] load balancing large farms?
> >
> >
> > This is somewhat inaccurate. I have several ServerIrons, as well as
> > several Arrowpoint CS800s. I can do fully customizable layer 7 health
> > checks where I define all the fun stuff I want ... it does
> > have to be an
> > HTTP request, but I can (and do) use layer 7 health checks. The only
> > caveat with Arrowpoint is that the response for a healthy
> > server must be
> > the same as they check a hash of the response .... ServerIron easily
> > checks for a customizable string in the response.
> >
> > For people like you with a need for non-HTTP health checks,
> > this can get
> > dicy, but for the most part, you're ok.
> >
> >
> > Troy Davis wrote:
> > >
> > > On Tue, 17 Oct 2000, Daniel Hagerty <hagIZZATlinnaean.org> wrote:
> > >
> > > > Because of this, I'm quite interested in hearing about what
> > > > options vendors offer for health checking and load balancing
> > > > decisions. Easy, push based methods of reporting
> > health/load to the
> > >
> > > We looked at 3 ASIC-based LBs - Foundry ServerIron,
> > Arrowpoint CS150,
> > > and Alteon AceDirector 3 - and only the Alteon did
> > customizable layer 7
> > > health checks where we could define a port to connect to,
> > the string to
> > > send, and the response to expect. The others would only do layer 4
> > > connections, nothing to verify that the server was actually
> > processing
> > > comands.
> > >
> > > We use protocols that nobody supported health checks on
> > (RTSP, HTTP with
> > > a required set of headers) and one of the known failure
> > cases we wanted
> > > to test for was a zombie server (still accepting connections but not
> > > doing anything).
> > >
> > > Lots of PC-based LBs (F5 BigIP, others) do scriptable layer 7
> > > verification, but it's scarce among ASIC switches.
> > >
> > > For your configuration, you could do balancing based on
> > least # of current
> > > connections, so that all servers would get 1 connection before any
> > > received 2 cons. If your non-standard protocol is UDP
> > based, verify that
> > > leastcons works with UDP.
> > >
> > > We've got a Cisco 2840G for port aggregation, then trunk
> > 4x100 Mbps ports
> > > to the LB, then the LB's gigE port connects to the router. Works
> > > nicely.
> > >
> > > What I really wanted was a 2 or 4 port gigE Alteon -- a
> > cheap version of
> > > the 8 port Alteon 180. We'd have used 1 port to plug the
> > upstream into
> > > and the other port for the aggregation switch. Ah well..
> > >
> > > Cheers,
> > >
> > > Troy
> >
> >
> > --
> >
> > Jeremy Noetzelman
> > Director of Network Architecture
> > PeopleLink
> >

each one has 3 gig blades and one 10/100 blade in addition to the
standard management modules.

they go downstream to Foundry BigIron 8k's.

approximate rule count is 500 or so, I don't have the latest number

-- 
 
Jeremy Noetzelman
Director of Network Architecture
PeopleLink



This archive was generated by hypermail 2b29 : Sun Nov 05 2000 - 15:09:07 EST