The Parallel Workload Analyzer Homepage
The workload analyzer was written by David Talby to analyze traces of empirical and synthetic workloads of supercomputers. The program and its source code are freely available under this license. The program works on traces in the Standard Workload Format (version 2.0), and can produce synthetic workloads in this format based on several distributions.
The current version of the program is 4.4, and its strongest points are:
- Three methods of testing for self-similarity for any variable are programmed: The variance-time plot, the pox plot, and the periodogram.
- One can easily filter workloads based on any combinations of conditions. It is then possible to remove filtered records from the logs, save them in a separate file, or receive statistics regarding the records that possess the checked property.
- Whenever a workload file is read, each record (job) is checked for all validity constraints required by the standard workload format. Invalid records are discarded.
- All modules of the program handle missing values.
- It is possible to graphically plot the distribution, aggregated distribution over time, auto-correlation and linear regression between any two variables.
- The ability not only to produce new synthetic workloads, but also to alter specific fields of existing ones by adding or multiplying them with a selected distribution.
- Batch analysis support for summary statistics, self-similarity tests, auto-correlation tests and per-month summary analysis.
- A simple and coherent graphical user interface.