Citation Context

...es. SystemML [7, 2] takes an approach similar to ours, but only considers optimization of execution, while Cumulon additionally considers hardware provisioning and configuration setting as well as bidding strategies for transient nodes. The root of this difference lies in Cumulon’s focus on the users’ perspective: while other systems optimize execution time on a given cluster, Cumulon goes a step further and asks what cluster would be most cost-effective (and how to bid for it) in the first place. Although automatic resource provisioning has been studied recently for general MapReduce systems [13, 23, 29], our focus on declaratively specified matrix-based programs leads to different challenges and opportunities in joint optimization of provisioning and execution that better exploits program semantics. There has also been a lot of related work from the HPC community on automatic optimization of matrixbased programs, though none has gone as far as automatically making provisioning and configuration decisions. Most closely related to Cumulon is the work on “semantic” optimization of MATLAB programs (i.e., at the level of linear algebra operators) [17]. At one point, MATLAB provided a chain-matrix...

Citation Context

...mization strategy to automatically determine the optimal job settings along the workflow for efficient execution and resource usage. Performance modeling framework. The performance modeling framework =-=[6, 11, 12, 5]-=- contains: (1) A platform performance model [6, 5] to predict the completion time of different MapReduce phases as a function of processed data. Specifically, we used a set of microbenchmarks to profi...

Citation Context

...nstrated at major conferences such as SIGMOD [48], SIGCOMM [26,29], NSDI [12,46], and COMSNETS [7], and experimentally deployed on testbeds such as PlanetLab [23,49], ORBIT wireless [7–9], Amazon EC2 =-=[41, 42, 47]-=-, and AT&T’s ShadowNet [11]. In the rest of this statement, I will summarize the research work in my group, broadly classified under different stages of the distributed systems development cycle: synt...