Abacus Project - The Diigo Meta page

• 02 Nov 07
  

  Bill H

  function_placement

  • While software engineering techniques (e.g., modularity and object orientation)  have given us the ability to partition applications into a set of interacting  objects, we do not yet have solid techniques for determining where in the  cluster each of these functions should run, and deployed systems continue to  rely on complex manual decisions made by programmers and system administrators
  • Optimal function placement in a cluster is difficult because the right answer is  usually, "it depends." Specifically, it depends on a variety of cluster  characteristics (e.g., communication bandwidth between nodes, relative processor  speeds among nodes) and workload characteristics (e.g., bytes moved among  functions, instructions executed by each function).
  • Some are basic hardware characteristics that only change when something fails or  is upgraded, and thus are relatively constant for a given system. Other  characteristics cannot be determined until application invocation time because  they depend on input parameters
  • Worst of all, many change at run-time due to phase changes in application  behavior or competition between concurrent applications over shared resources

  2 more annotations...

Would you like to comment?

Join Diigo for a free account, or sign in if you are already a member.