CALIENT Technologies

Move the light, not the fiber

Solving The Elephant-Mouse Data Center Data Flow Challenge

By Daniel Tardent

When elephants stampede in the wild, the ground shakes, the dust flies and everything in the way moves – or else.

When elephant data flows stampede in your data center its just the opposite – there’s no noise, no dust and, unfortunately, no movement – as all other packets get stuck behind these long data flows.

Courtesy of Flickr User YoTuT

Courtesy of Flickr User YoTuT

I’m in Austin, TX, this week at the Lightwave Optical Innovation Summit to speak about elephant and mouse data flows and how data center network managers can build networks with the performance for both.

Mouse data flows are emails, web pages, data requests or any other short-lived data flow.  Elephant data flows, on the other hand, are persistent flows such as VM migrations, data migrations, MapReduce and other application flows that impact network bandwidth for minutes or hours or more.

Many networks are sized to handle peak traffic only in bursts, because of the high cost and power consumption involved in of building a non-oversubscribed network. Elephant flows have always been with us, but they are now increasing in frequency and in length due to growth in virtualization and big data.

Building a Dedicated Lane for Elephant Flows

It’s a problem when you consider that there is so much east-west traffic in the data center (between 75% and 95%), which tends to require very low latency. It’s at a point where a lot of data center managers I speak with say these flows are significantly impacting their data center performance.

Certainly, building a non-blocking packet network is one way to deal with the issue of elephant traffic, but what I’ll be telling the folks in Austin is about how to build an optical circuit switching overlay network – what we call a hybrid packet-optical circuit data center network – that provides a low-latency channel dedicated to persistent data flows.

I’ll give an example of how this network works in what I call a “super cluster” configuration, that is four clusters each with 48 racks that each consist of a top of rack switch with two 10G and six 40G uplink.  I’ll also talk about aligning this approach with software-defined networking (SDN) controllers.

Cost, Power Advantages

And, I’ll also touch on both the cost and power consumption advantages of the hybrid packet-optical circuit networking approach.  These issues are almost as important as the throughput specs, and they are a significant advantage for the hybrid approach that only gets better as the bandwidth increases – a function of the fact that optical circuit switching can support data rates between 1Gbps and 400 Gbps with no need to swap out optics or interfaces.

Here’s a summary of those advantages for 40GB hybrid network vs. non-oversubscribed all-packet network:

  • 10x lower capital expense
  • 150x lower power
  • 11x – 26x lower total cost of ownership

I’m looking forward to the feedback from those that attend my session.  If you want more information on this, just drop me an email and I will send you my slide deck for review.

Comments are closed.