Which QoS approach for this scenario?

Unanswered Question
Aug 20th, 2008

Imagine I have Router3825Headquarter connected to MY_MPLS network.

Then I have 3 2851 partner offices connected to MY_MPLS network. Servers are located on Headquarters. All sites including headquarter connected via 9Mbps circuit.

Management is asking:

What happens if one branch office starts sending very large files and overutilize the headquarter circuit.

In this case, where do I need to implement QoS mechanisms to proper bursts and provide adequate bandwidth for other partners who also connect to my headquarter?

I have this problem too.
0 votes
  • 1
  • 2
  • 3
  • 4
  • 5
Overall Rating: 0 (0 ratings)
Loading.
Edison Ortiz Wed, 08/20/2008 - 11:31

If you want to ensure proper allocation of bandwidth in all locations, then QoS must be configured in all locations.

__

Edison.

tdrais Wed, 08/20/2008 - 13:19

I am not a big fan of MPLS for exactly these issues.

So if you want to attempt it by yourself without you MPLS vendor.

At each of the spoke sites limit the outbound traffic to 1/3 the bandwidth. This would ensure you never exceed the total bandwidth. But of course you waste 2/3 your bandwidth. To deal with the spoke to spoke traffic you would have to do this for your main site also but with 3 shapers.

So I can assume this is not acceptable. The only real way to do this is to get the MPLS provider to do it for you. It will all depend on how they do their QoS and what they will allow you to do.

So lets assume there are 2 classes the MPLS provider allows. class 1 no drop. class 2 drop on congestion.

So now at each remote site you would put in the same policers as the first example. You would police all traffic to 1/3 the bandwidth and mark the conform traffic to 1 and the exceed traffic to 2. You would still send all the traffic

With this method if all the remote sites were attempting to send at max rate to the main site each would get 1/3 the bandwidth if any site did not use all his class 1 other sites could send class 2 traffic during this time.

This is a overly simplistic example but all are based on using policers to mark traffic and those markings are based on what types of classes the provider has.

You have to read your contract very carefully to even see if you could transfer traffic at max rate from just one spoke site to the main site. You may have a best effort contract which means they could drop 99.9% and still be ok.

news2010a Wed, 08/20/2008 - 21:03

Well, it sounds yes, I will have to explore this with MPLS provider then.

I think this idea of implementing 1/3 of the bandwidth at each site is really not very practical.

Joseph W. Doherty Wed, 08/20/2008 - 15:14

"What happens if one branch office starts sending very large files and overutilize the headquarter circuit."

Default behavior is likely to be much like if the four sites were inconnected with a LAN switch without any QoS, i.e. probably FIFO with tail drop.

"In this case, where do I need to implement QoS mechanisms to proper bursts and provide adequate bandwidth for other partners who also connect to my headquarter? "

Where? If single queue FIFO isn't acceptable, QoS should be enabled wherever congestion causes issues. The two most likely initial spots are at both ends of your MPLS links, both egress (hopefully under your control) and ingress (which the MPLS vendor controls).

As Tim correctly notes, you can avoid HQ MPLS egress congestion if you keep the all your remotes sites from being able to consume more than the HQ's link bandwidth. As he further notes, this could be done by limiting each remote site to 1/3 the HQ link's bandwidth, but any combination that sums to the HQ link's bandwidth works too, e.g. 1 Mbps, 2 Mbps and 6 Mbps for the three remote sites.

Otherwise, you'll need to determine what QoS the MPLS offers, if any, and work it to your needs. MPLS vendors often differ in their QoS support and may even offer different QoS configurations (sometimes at extra cost).

Steve Lyons Wed, 08/20/2008 - 16:07

The difficulty in this scenario is trying to control traffic coming into your HQ site. The two easiest ways to avoid overruns and packet loss is to either set the connection speed of HQ to equal all the remote sites connecting to it or throttle the remote sites so the total aggregate of the remote sites does not overrun the HQ site.

Steve Lyons - Cisco

Actions

This Discussion