L4 vs L3 LACP Load Balancing - What's the Use Case?

[Skaffen]

Can someone explain when you'd chose L4 over L3 load balancing for an LACP trunk? The manuals say how you do it, but not why you'd use it.

L3 is the default, but I'm curious as to what the benefits of L4 are, and when you'd use it.

[parnassus]

L4 hash, with respect to L3 hash, provides an hashing mechanism that may provide a better redistribution of the outgoing traffic across the physical links of a LAG resulting with an even distribution (so a better balancing) exactly because it adds TCP/UDP ports into its hashing calculations and that may introduce more variations (if Source and/or Destination TCP/UDP ports are effectively used and varies on traffic flows leaving the Trunk), those variations are exactly needed to (re)distribute the traffic better than it can happen using only L3 hashing: clearly distribution efficiency depends not only by the hashing algorithm but also by the type of traffic really flowing through the considered Trunk.

[Skaffen]

Thanks for the reply.

So from your description, and the manual saying that L4 hashing will fall back to L3 if there's no port info, L4 hashing would be the best option as long as the switches have ample resources for the additional load?

[parnassus]

What Switch(es) exactly are you referring to?

[Skaffen]

 A 5412R zl2 as a collapsed core, and six stacks of four 2920s at the edge. 2 X 10 Gb LACPed to each stack.

[parnassus]

I can't prove what I'm writing (no easy way to understand [*] what are L3/L4 hashing capabilities and in which way L3/L4 hashing computations were implemented into each Switch series) but, considering that the Aruba 5400R zl2 Series supports up to 144 LAG (Ports Trunks), each one with up to 8 physical links (Ports) per Trunk, L3/L4 hashing computations shouldn't be really an issue [**]; same can be said about Aruba 2920 series (backplane stacked or not they are).

[*] Would be interesting to know...

[**] Aruba 5400R zl2: HPE declares a 2 Tbps crossbar switching fabric capability which provides an intra-module and inter-module switching throughput of 785.7 Mpps on the purpose-built ProVision ASICs.

[Skaffen]

Yeah, I'm pretty sure the 5412R will handle it without batting an eyelid.  We'll only be using 21 LAGs initially.  I might run some benchmarks on L3 vs L4 load balancing just for interests sake to see if there's any noticeable increase in load on the core.

I priced up a stacked core, but the modular option was less expensive and has twice the throughput and more flexibility. 

Thanks for your input.  The switches haven't arrived yet so I'm working my way through the various manuals to rough out a config, and asking questions for the features I'm not familiar with.

[parnassus]

Skaffen wrote: We'll only be using 21 LAGs initially.

Twenty one LAGs? ...are you including those ones dedicated to Server hosts in the count too? ...since you wrote you have just six uplinks to your six Aruba 2920 backplane stacked's pairs.

Skaffen wrote: I priced up a stacked core, but the modular option was less expensive and has twice the throughput and more flexibility.

 Just curious...with "stacked core" were you referring to a dual Aruba 5400R zl2 configured to be deployed with VSF (Virtual Switching Framework) or what else? ...in both cases (two units front-plane stacked or just one unit alone) the used Aruba 5400R zl2 is "modular"...so I'm asking.

[Tobias_IT]

Hi, what about the switch JH390A?

L3/L4 Load Sharing Balancing Mode on HPE 5940 - JH390A.  it is possible?

Thanks,

[parnassus]

Yes, please have a look at the latest pulished HPE FlexFabric 5940 Switch Series - Layer 2 - LAN Switching Configuration Guide to discover how to set globally or punctually the load balancing algorithm.