TCP over IP Anycast - Pipe dream or Reality?

The talk will focus on how to route our end users to the closest location serving content – i.e. to the closest PoP. Traditionally LinkedIn used geo-location base …

The talk will focus on how to route our end users to the closest location serving content – i.e. to the closest PoP. Traditionally LinkedIn used geo-location based load balancing (with help of DNS) but there are challenging areas with this approach that lead to bad performance for the end user and operational challenges for the LinkedIn site teams.

1. Sub-optimal routing due to the fact that geographical load balancing makes DNS mapping decisions based on an end users' name server IP as opposed to the client.
2. Geographical mapping of IPs to latitude/longitude and city-level mapping is not 100% accurate nor related to fiber and internet connectivity map.
3. Operational complexity. As we grow the number of PoPs, it will be increasingly difficult to scale the use of geographical load balancing. The answer to the above challenge, was TCP anycast: Anycast provides a distributed service via routing. Based on anycast routing, packet will arrive to the closest node depending on the location of source and hop-by-hop routing decisions. In short the talk will discuss about how we improved our anycast implementation using bgp and how we measure the success rate: How LinkedIn used TCP anycast and RUM to drive optimizations and make the site faster. Measurements using RUM is demonstrated in the following presentation: http://www.slideshare.net/rmaheshw/velocity-2015-pops-and-rum blog post: https://engineering.linkedin.com/network-performance/tcp-over-ip-anycast-pipe-dream-or-reality Presenters: Shawn Zandi and Ritesh Maheshwari