======================================================================== UCit Predict-IT ======================================================================== Information on Predict-IT is available at: https://www.ucit.fr/en/predict-it/ --------------------------------------------------------------------- Release 1.5 - July 9, 2020 --------------------------------------------------------------------- NEW: - Added support for "data enhancers". You can provide csv files that contains data to be added to the job scheduler parsed data, given a pivot column and a column of new values or a Python file to add any type column or modify the content of the job scheduler’s data. New features are then automatically used for training. This allows for example the addition of application specific metrics. - Added 15 new features related to target statistis using data enhancers. - Predict-IT now trains the selected model once more with all data available, instead of keeping a model trained only on the train dataset (thus missing the jobs in the test dataset). This allows for better accuracies in production. - Added the possibility to launch multiple Predict-IT in parallel with different configurations. This allows to build specific predictors (e.g., per application), or to test multiple configurations. ENHANCEMENTS: - Added new configuration options for model features and bins selection. - Added support for slurm WCKEYS column - Added support for automatic batch retrieval as smaller requests for big sacct call BUG FIXES: - Fixed tracking script: tracking dates were not updated everytime - Fixed Slurm 19 error while extracting data - Fixed an error that poped when the server configuration file was specified on the command line, and the default path did not exist. - Various bug fixes related to missing data in the input data --------------------------------------------------------------------- Release 1.4 - March 11, 2020 --------------------------------------------------------------------- CHANGES: - Added filter on input dataset. ENHANCEMENTS: - Added last training status on metrics page. - Added server parameters in help page. - Added a way to run multiple PIT configurations using server.conf files with specific port, host and dirName. - Added comparison between previous and newly trained models. - Added a way to turn on / off the use of balanced bins directly within server.conf. BUG FIXES: - Error when computing F1-score for unique class (e.g. after filtering). --------------------------------------------------------------------- Release 1.3 - December 17, 2019 --------------------------------------------------------------------- CHANGES: - Added wait time and time to result prediction. - Added tracking-script functionality to collect jobs since last X hours, call for predictions on each of them and visualize them. - Added option in server.conf to save training "test" metrics as csv files (Obs, Pred, Confidence). - Added option in server.conf to define the total CPU number of the cluster. ENHANCEMENTS: - Added 2 RMSE calculations in metrics. - Added unweighted accuracy gauge in metrics to investigate global accuracy without taking into account the support for each class. - Redesigned help page. BUG FIXES: - BAD_REQUEST when requesting prediction using predictit.client. - Missing column when handling SGE. - Fixing Comparison page. --------------------------------------------------------------------- Release 1.2 - July 26, 2019 --------------------------------------------------------------------- CHANGES: - Added AdaBoost in the list of available algorithms. - New EULA. - New logo added. ENHANCEMENTS: - Script extractSlurmData.sh now has a -e/--end option to specify the end time. - To extract data from the job scheduler, you can now directly use bin/extractData , instead of /jobschedulers/extractData.sh - It is now possible to ask for the last N days of data instead of having to define a specific starttime. BUG FIXES: - Many small bug fixes and typos have been corrected in all the in the job scheduler parsers. - Torque/PBS: better handling of non utf-8 characters. - SLURM: Extract job scheduler data in local time instead of UTC. --------------------------------------------------------------------- Release 1.1 - August 28, 2018 --------------------------------------------------------------------- CHANGES: - Added memory consumption (Max RSS) prediction. - Predictions now come with a confidence level (from 0 to 1). - init.d scripts. ENHANCEMENTS: - Better support for Slurm <15. - Faster parsing of Slurm accounting logs. - Metrics web pages now display a global indicator showing the accuracy of the model, along with a graph showing the evolution of this metric through time. - New installation directory organization to ease update procedures. BUG FIXES: - Accessing the metrics web pages when the models haven't been trained do not display a JSON error anymore. - Fixed various graphical bugs in the metrics web pages. - Fixed bug that happened when cv=False. - Correctly parse job steps in Slurm logs. --------------------------------------------------------------------- Release 1.0 - February 6, 2018 --------------------------------------------------------------------- First version of Predict-IT. See documentation for a complete description of Predict-IT features.