BGP Blackmagic: Load Balancing in "The Cloud"

There are all kinds of ways to load balance a cluster of servers. Some operators purchase big devices from Cisco or F5 or Foundry and some people prefer to use specific software tools like pound and perlbal. But these tools have limits—they are often only effective on a local network. When it comes time to balance your service across an array of datacenters and provide failover things can get complicated. DNS load balancing only gets you so far and is often slow to deal with failover. Some folks just give up and outsource to CDNs like Akamai or Limelight. But that doesn’t always work with complex dynamic content. So how do you scale out? This is a conference for operators and architects of large systems.

Let’s talk about designing scalable networks and architectures that span multiple datacenters. Oh, and let’s do it with the “cheap but reliable” methodology. Everything I am going to talk about is doable with open source software and commodity hardware. Also, everything I will discuss can be done in stages, as you grow, I mean… SCALE UP. :-)

Sections of presentation:

• A quick review of how big networks exchange packets.

A BRIEF overview of how networks communicate and interior and exterior routing protcols and their uses. Also a slide on numbering related to IP addresses, allocations, AS numbers, and ARIN/IANA, etc. Also a quick explanation of the difference between a router, switch, load balancer, and server.

• Smashing the (OSI) stack and load balancing at Layer 3.

An overview of what a “layer 3” routing protocol is, like OSPF or BGP, as opposed to something deeper like squid or perlbal, and how they can be as effective or more effective both locally and globally.

• A BGP Primer—prefixes and routing. Also introduce the often-discussed-but-rarely-understood term of Anycast.

• An example of local (intra-area) load balancing in a single datacenter with BGP Anycast.

• An example of global (inter-area) load balancing with three datacenters with BGP Anycast.

• An example (live!?) of failover between sites or servers using BGP.

• Discussion of caveats with UDP and TCP Anycast between sites.

• Discussion of tools, techniques, and resources for deploying these technologies.

• 10-15 minutes of questions… with BGP there are always questions. :-)