Doug Gourlay of Arista Networks sat down for an interview with SiliconANGLE’s John Furrier and Wikibon’s Dave Vellante, providing a glimpse into the networking layer of a Hadoop deployment.
After a brief talk about the growth his company has seen recently, Gourlay discussed the differences between the traditional datacenter network, and the one used to transfer data within a Hadoop cluster. He explained that Hadoop distributes the data across a large number of nodes and replicates it for data protection purposes. This process takes a Layer 3 Hadoop-aware file system, compared to the “fat and flat” networks needed for virtualization. “You can route it,” Gourlay added to sum up the requirements of the big data analytics engine.
Furrier went on to ask what some of the disruptions the networking veteran believes Hadoop will drive among traditional players, and Gourlay responded by highlighting the storage industry first and foremost. Hadoop, for all its open aspects, eliminates the need for multi-million storage arrays in favor of more economic solutions, according him, which means storage vendors are going to have to adapt. The same can be said of the companies that are relying on their analytics infrastructure, if they are looking to pursue the advantages of consolidating their clouds.
Speaking from experience, Gourlay stressed that organizations should be focusing on squeezing in more servers in a smaller footprint – operate at scale, and bring what would be several small datacenters-worth of servers under one room in order to cut overhead costs. This approach requires a change to the existing IT model, which also encompasses a number of other things.
Arista’s VP of marketing added that companies should be centralizing the IT department as well, and extending the role of their engineers to stretch across a number of areas. A lot of companies having already opened up to this notion, such as Facebook, and the demand for data scientists – professionals who have a skill-set covering both both programming and analytics – is a strong indicator of that.
Latest posts by Maria Deutscher (see all)
- LinkedIn previews upcoming enhancements for Apache Kafka - September 2, 2015
- Syncsort open-sources Spark connector for mainframes - September 1, 2015
- SAP debuts an in-memory query engine for Spark in renewed IoT push - September 1, 2015