Load balancing on 2 MPLS p2p connection

[sreeraj.murali]

Hi Experts,

We have 2 MPLS P2P circuits between 2 offices. The circuits are used to cater replication traffic between offices.

Please advice best feasible option to effectively use both the circuits working together-load balancing, load sharing, bundling etc.

                |------mpls 1-------|

Office A--|                                  |--Office B

                |------mpls 2-------|

Thanks in advance

Sreeraj Murali

[Giuseppe Larosa]

Hello @sreeraj.murali,

IF your routers can see each other on the two MPLS circuits without  Service provider devices in the middle ( not an MPLS L3 VPN service) you can use OSPF or EIGRP and you can implement ECMP Equal Cost Multi Path

This is flow based , all packets of a given flow ( specific IP SA , IP DA) will use a single link per direction

Hope to help

Giuseppe

[sreeraj.murali]

Hi,

The MPLS circuit is from 2 different providers TATA and Airtel. This is provider based MPLS service.

What we can do in this case ? Please advice.

Thanks and Regards

Sreeraj Murali

[Giuseppe Larosa]

Hello Sreeraj Murali,

what do you exactly mean with "This is provider based MPLS service." ?

in L3 MPLS VPN a peer model is used your devices will exchange routing information with Provider Edge PE nodes.

In a L2 MPLS service your routers can talk to each other directly as the service provide a point to point OSI layer 2 transport of frames.

Two different Service Providers is good for redundancy and fault tolerance

So you need to clarify if you have an MPLS L3 VPN service between two sites or you have bought two L2 point to point transport services from the two providers.

In the second case you can easily implement ECMP using OSPF or EIGRP directly on your routers they share a common IP subnet on each link.

Reading again your first post  in this thread I think you are in case 2).

To verify this just check that you are free to choice the IP addresses to be used on each MPLS link and that your routers are direct neighbors at IP level ( OSI layer 3)

TATA L2 Link :    10.1.1.1/30    on R1       10.1.1.2 on R2

Airtel L2 link 10.1.1.5/30 on R1 10.1.1.6 on R2

With OSPF for example

R1:

router ospf 10

router-id 1.1.1.1

network 10.1.1.0 0.0.0.3 area 0

network 10.1.1.4 0.0.0.3 area 0

network 10.100.0.0  0.0.255.255 area 0

! note here 10.100.0.0/16 represent all subnets in site 1

R2:

router ospf 10

router-id 2.2.2.2

network 10.1.1.0 0.0.0.3 area 0

network 10.1.1.4 0.0.0.3 area 0

network 10.200.0.0  0.0.255.255 area 0

! note here 10.200.0.0/16 represent all subnets in site 2

Hope to help

Giuseppe

[Richard Burts]

I would like to understand better this statement "The circuits are used to cater replication traffic between offices." Is this replication traffic a single host at one site to a single host at the other site? Or is this multiple hosts at one site or multiple hosts at the other site?

If there are multiple hosts involved in this at one or both sites then the ECMP suggestion with dynamic routing protocols may be a good solution. But be careful if the replication is a single host to a single host. I would like to share an experience that I had with a customer. This customer had a main site and another site. They were sending replication data from a server at one site to a device at the other site. The time required to send the data was getting long. So they implemented a second connection and to effectively utilize both connections they implemented per packet load sharing (rather than the default per flow sharing). The day that they implemented the second connection they eagerly waited for the results of the replication. To their major surprise the replication process took significantly longer than was normal. As we investigated what was happening we realized that with per packet load sharing there were numerous occurrences of out of order packet delivery. And the error detection and recovery for out of order packets was making the replication much slower. Their solution was to route most traffic over one link and to route the replication traffic over the second link. Perhaps a solution like that might work for you?

[Joseph W. Doherty]

BTW, what Rick notes about out-of-order packets causing an adverse performance impact is generally the principle reason why using per packet load sharing is not recommended.  (NB: something like a hardware MUX or PPP will place packets back into order before forwarding them when they cross multiple links.)

It's possible, on some hosts, to adjust the TCP stack (on sending host) to be more "tolerant" of out-of-order packets, which if set correctly, would allow per packet load sharing to work very well.  However, making such changes a) also degrades TCP reacting to some issues which its default settings are defined to detect, b) there's the issue of adjusting hosts and c) optimally "tuning" the parameters.

Oh, if the replication is using TCP, there might also be merit it "tuning" (receiving) host's RWIN for the BDP of WAN.  This help insures optimal TCP transmission rate.

[balaji.bandi]

Is this circuit the same capacity or a different capacity?  is this from the same Service Provider?

you can do as suggested with OSPF. or you can do different traffic using different circuits and failover if one fails to other.

[MHM Cisco World]

friend divide the subnet in to two half one through one SP and other half through other SP. 
Q. from CE-PE what protocol you run ?

[Joseph W. Doherty]

If both sites have Cisco routers that support PfR, that might be one of your best options as it can dynamically load balance.  (This assumes enough of a flow mix that you and effectively balance both links.)