Cloud computing promises flexibility and high performance for users and lower costs for operators. Despite a shift from dedicated to shared clusters, most datacenters still operate at low utilization for several reasons, including exaggerated resource reservations, workload interference and platform heterogeneity.
In this talk from Twitter's recent open source #CONF, Christina Delimitrou will present an analysis of utilization at Twitter and discuss a few techniques to improve it. First, for long-running applications, the cluster manager can estimate near-future load changes and adjust resource allocations accordingly. Second, given the difficulty users experience in determining appropriate resource allocations, she will describe a different interface with the cluster manager that requires the users to specify performance goals as opposed to raw resources. This approach enables significantly higher cluster utilization without degrading per-application performance.
Join the conversation at
http://twitter.com/university
Show less
