Process large amount of data and to implement complex data analyses using Spark. The dataset has been made available by Google. It includes data about a cluster of 12500 machines, and the activity on this cluster during 29 days.
To start the Spark application, please run the script run-app.sh.
The following questions are answered in this lab:
• What is the distribution of the machines according to their CPU capacity?
• What is the percentage of computational power lost due to maintenance (a machine went offline and reconnected later)?
• What is the distribution of the number of jobs/tasks per scheduling class?
• Do tasks with a low scheduling class have a higher probability of being evicted?
• In general, do tasks from the same job run on the same machine?
• Are the tasks that request the more resources the one that consume the more resources?
• Can we observe correlations between peaks of high resource consumption on some machines and task eviction events?
• Do tasks having the higher priority require more resources?
• What are hardware specifications of machines on which different priority tasks have/haven't successfully run?
The solution is compared from different points of view (performance, ease of use,etc.) with different technical solutions to process the data:
• Compare with Spark Dataframe.
• Compare the use of Spark to the use of a non-parallel Python data analysis library such us Pandas.
Our Spark application is deployed in a real distributed environment (GCP Dataproc cluster) to conduct some performance evaluation in this context.