Inside the Social Network’s (Datacenter) Network By: Austin Christian Paper written by: Arjun Roy, Hongyi Zheng, Jasmeet Bagga, George Porter and Alex Snoeren
Main Paper Discussions A look inside Facebook’s datacenter to see how network traffic is managed. Discuss contrasting locality, stability, and predictability of network traffic in Facebook’s datacenter. Comment on the implications of Facebook’s network traffic(locality, stability, and predictability) for network architecture, traffic engineering and switch design.
Facebook’s Network Topology Machines organized into racks and connected to top-of-rack switches (RSW) @ 10Gbps Ethernet links. Each RSW connected via 10Gbps links to four aggregation switches called cluster switches(CSW). CSWs connected by another set of aggregation switches called Fat Cats (FC).
Datacenter server jobs A unique aspect about Facebook’s datacenter is servers do one job and do not deviate from that task. Web Servers (Web) serve web traffic MySQL servers (DB) store user data Cache Servers (Cache) handle temp storage for query results, handle cache coherency, serve most read requests Hadoop servers (Hadoop) handle offline analysis and data mining Multifeed servers (MF) assemble news feeds Racks do not contain more than one type of server( ie A rack consisting of only Web servers)
Locality and Stability Web Server Hadoop Server Cache Leader Cache Follower
Locality and Stability Cont. What is observed from the graphs: - Hadoop Server is only server out of the four which is rack-local (traffic stays within cluster and not very stable over the 2min interval). (Note the cyan bars). - Web server, Cache Leader, and Cache Follower display constant stability over the 2min interval Most of the web server traffic stays within the cluster and majority of it is destined to the cache. (Note green and blue bars). Due to cache followers interact with a majority of hosts in cluster and communicates with over most of the web servers. Thus the high intra- cluster traffic (blue bars) The cache leader however engage in primarily inter and intra- datacenter traffic, since cache leader communicates with multiple servers on a large scale (thus the red and green bars).