Dynamic Balancing and Hyperflow real world experience

[Albin]

Hi!

We are thinking about going for Dynamic balancing with Hyperflow for our VSX and Security gateways. Today they are configured with manual affinity settings for each VS. 

I see some other posts about this but with small amount of input of real world experiences of dynamic balancing. Most of them are also two years old. Hoping for some new input if there are any caveats, or if it behaves well?

We have 26000 with VSX & 28000/7000 with Security Gateway.

[AmitShmuel]

Hi,

Starting with a positive tone, over the years I've seen many customers who have benefitted greatly from these features (Dynamic Balancing obviously has more years in the field).

For Dynamic Balancing, the optimal case would be an out of the box CPU balance between SND cores, and FWK.

An example of this can be seen below, taken from a real customer (Left: before enabling the feature, SNDs are heavily utilized. Right: after enabling the feature, SNDs and FWKs are perfectly balanced).

For HyperFlow, there are two main benefits:

1. Elephant Flows throughput boost
2. CPU spikes reduction (A decrease in the CPU utilization of the FW worker handling the Elephant Flow - this depends on the possible throughput of the elephant flow, if we can process more of it, the CPU utilization will remain high, even if we add more HyperFlow workers)

Here is an example of HyperFlow customers showing the CPU average utilization of FW workers, while they process an elephant flow, with and without HyperFlow. (Left: synchronous processing, without HyperFlow. Right: asynchronous processing, with HyperFlow. again, the decrease depends on what I've described above).

Since R81.20 is HyperFlow's first release, I would recommend using the latest JHF take, as several fixes may have already been published.

Thanks,
Amit