Inside the Social Network’s (Datacenter) Network – Public Review

This paper presents a first look into the traffic characteristics of the workloads in datacenters operated by Facebook. In doing so, the paper extends the publicly available knowledge-base on production workloads. As well, it offers some novel and interesting measurements. First, the Hadoop clusters operated at Facebook exhibit substantially smaller amounts of rack locality than previously reported. That is, more than 80% of the traffic crosses racks. Second, in spite of the above lack of locality, the overall network link utilization is quite small– an average of less than 10% on all potential bottlenecks. Links that support Hadoop traffic are much more likely to have a higher load relative to other links in the datacenter. And, diurnal changes in request load from users accounts for only a modest variation in network traffic (about 2×). Third, the other dominant application at Facebook is memcached-style request-response workload which comprises primarily of small packets and has some specific traffic patterns. Finally, the paper posits lack-of-stability-in-the-trafficmatrix, poor-predictability-of-heavy-hitters and goodapplication-level-load-balance as reasons that make fine grained traffic engineering not needed or not easily doable.