Source: Dissertation Abstracts International, Volume: 75-05(E), Section: B.

Adviser: Michael J. Fischer.

Cloud computing is playing a vital role for processing big data. The infrastructure is built on top of large-scale clusters of commodity machines. It is very challenging to properly manage the hardware resources in order to utilize them effectively and to cope with the inevitable failures that will arise with such a large collection of hardware. In this thesis, task assignment and checkpoint placement for cloud computing infrastructure are studied.

As data locality is critical in determining the cost of running a task on a specific machine, how tasks are assigned to machines has a big impact on job completion time. An idealized abstract model is presented for a popular cloud computing platform called Hadoop. Although Hadoop task assignment (HTA) is NP -hard, an algorithm is presented with only an additive approximation gap. Connection is established between the HTA problem and the minimum makespan scheduling problem under the restricted assignment model. A new competitive ratio bound for the online GREEDY algorithm is obtained.

Checkpoints allow recovery of long-running jobs from failures. Checkpoints themselves might fail. The effect of checkpoint failures on job completion time is investigated. The sum of task success probability and checkpoint reliability greatly affects job completion time. When possible checkpoint placements are constrained, retaining only the most recent O(log n) possible checkpoints has at most a constant factor penalty. When task failures follow the Poisson distribution, two symmetries for non-equidistant placements are proved and a first order approximation to optimum placement interval is generalized.