From: Ryan Morris (ryan@egate.net)
Date: Thu Dec 06 2007 - 12:12:48 ART
Hi Rudi,
It turned out to be pretty hard to test in production. I put a traffic
generator on a generated lots of critical data. Then I started a huge
Bulk data transfer, while on the phone (IP) to a tester in the remote
office. It appeared that the QoS config pushed the two types of traffic
back a bit, but they both ate up more than their queues (which makes
sense, since at the time there was quite a bit of free bandwidth in the
other queues). Also it's not instantaneous... I don't know how long it
would take for the traffic to balance out if you had traffic from all of
your queues on the wire.
I think next I need to take my traffic generator to the lab and create a
simple QoS config. Say, FTP traffic gets 25% and all other traffic
(default) gets 75%. Then I should be able to start a large FTP file
transfer and at the same time generate lots of http traffic, and see the
traffic (over time) be limited the way it is configured.
In production, I don't see much difference in our traffic patterns. I
think it's because we usually see a single protocol cause congestion,
while the other protocols are within their queue limits. It seems like
the router allows the single protocol to suck up all the bandwidth
available. I haven't seen a case where I can definitely say the router
adjusted the bandwidth of a misbehaving file transfer.
Ryan Morris
CCIE #18953
On Thu, 6 Dec 2007, Rudi Sutiono wrote:
> Hi Ryan,
>
> Have you test your scenario? How is the result?
> I have test my scenario, but i don't see significant different when i apply QoS and when i bypass the QoS. Can you give me sample configuration of CBWFQ and how to test the configuration to make sure it works well? Thanks...
>
>
> Regards,
>
> Rudi S
>
> -----Original Message-----
> From: Ryan Morris [mailto:ryan@egate.net]
> Sent: Wednesday, October 31, 2007 6:26 PM
> To: ccielab@groupstudy.com
> Cc: Rudi Sutiono
> Subject: Re: Need help for QoS
>
> Hi Rudi,
>
> I think two things could be at play here:
>
> First, if you're overloading all the queues you'll get drops no matter
> what your QoS policy is, which would cause your TCP flows to back off.
> This may be reducing the bandwidth used by your critical apps.
>
> Second, the bandwidth will be dynamically allocated. If your voice calls
> are using only 10% of the bandwidth, and your critical traffic 5%, the
> algorithm will allow your best effort traffic to take up the rest of the
> 85% of the available bandwidth. This is by design.
>
> How are you testing the critical data? If you have a traffic generator,
> try putting 25% voice, 50% critical and the rest best effort. The extra
> voice will be policed, and the rest should be delayed to get the ratio
> you've configured. I haven't tried this test yet with my scenario. I
> hope to do this kind of test early next week. I'll let you know how it
> goes.
>
> Regards,
>
> Ryan Morris
> CCIE #18953
>
> On Wed, 31 Oct 2007, Rudi Sutiono wrote:
>
> > Hi everyone,
> >
> > Now I'm developing QoS for my customer and I have some problems regarding QoS
> > that make me confuse, could you guys help me with this problems.
> > Current condition: My customer has 2 site offices (Jakarta and Surabaya), the
> > interconnection using 10Mbps WAN link, I want to do QoS for the WAN link with
> > this breakdown:
> > * Voice: 20% priority queue (to cover all the VoIP calls)
> > * Critical: 40% use for database application (Oracle), mail
> > application (Exchange), and some critical applications
> > * Best Effort: Everything else
> > With this current breakdown, I want to make sure that voice & critical
> > application still get best service when the WAN link congested with other's
> > traffic. To test this condition, I make the WAN link congested using WAN
> > killer application, all of the 10Mbps bandwidth used by best-effort class (use
> > "show policy-map" command). After that, I try to do large database replication
> > for Oracle (critical class) hoping that critical class will get 40% of the
> > bandwidth. But it takes very long time to map fairly between the classes,
> > because after 15 minutes transfer, the critical class only gets 5% of the
> > 10Mbps bandwidth (use "show policy-map" command). The questions are:
> > 1. Is this the normal QoS behavior or not?
> > 2. How long the time QoS takes to map fairly between the classes?
> > Because if it takes too long time, will affect the voice and critical class
> > performance.
> > 3. What is the best QoS implementation for WAN link and MPLS? Can you
> > guys give me details of QoS implementation for WAN link and MPLS link (also
> > the configuration)?
> >
> > Any thoughts or real world experience would be appreciated.
> >
> >
> > Best Regards,
> >
> >
> > Rudi Sutiono
> >
> > _______________________________________________________________________
> > Subscription information may be found at:
> > http://www.groupstudy.com/list/CCIELab.html
>
> _______________________________________________________________________
> Subscription information may be found at:
> http://www.groupstudy.com/list/CCIELab.html
This archive was generated by hypermail 2.1.4 : Tue Jan 01 2008 - 12:04:29 ARST