High performance shared state schedulers

Sammanfattning: Large organizations and research institutes store a huge volume of data nowadays.In order to gain any valuable insights distributed processing frameworks over acluster of computers are needed. Apache Hadoop is the prominent framework fordistributed storage and data processing. At SICS Swedish ICT we are building Hops, a new distribution of Apache Hadoop relying on a distributed, highly available MySQL Cluster NDB to improve performance. Hops-YARN is the resource management framework of Hops which introduces distributed resource management, load balancing the tracking of resources in a cluster. In Hops-YARN we make heavy usage of the back-end database storing all the resource manager metadata and incoming RPCs to provide high fault tolerance and very short recovery time. This project aims in optimizing the mechanisms used for persisting metadata in NDB both in terms of transactional commit time but also in terms of pre-processing them. Under no condition should the in-memory RM state diverge from the state stored in NDB. With these goals in mind several solutions were examined that improved the performance of the system, making Hops-YARN comparable to Apache YARN with the extra benefits of high-fault tolerance and short recovery time. The solutions proposed in this thesis project enhance the pure commit time of a transaction to the MySQL Cluster and the pre-processing and parallelism of our Transaction Manager. The results indicate that the performance of Hops increased dramatically, utilizing more resources on a cluster with thousands of machines. Increasing the cluster utilization by a few percentages can save organizations a big amount of money.