Thursday, March 16, 2023
MLIT Room 310
E.I. Aleksandrov

The Event Picking Service

The main use case for the ATLAS EventIndex is event picking, i.e. finding one or more events in the many billion events generated by the ATLAS experiment, stored in several million files, and extracting them. One of the latest addition to the EventIndex project infrastructure is the development of a new Event Picking Service, in order to automate the process of event picking for cases when the number of requested data is large.
One example of massive event picking in 2019 was the “γγ→WW” analysis. This analysis required the extraction of 50k events in RAW format out of the 18 billion events in Run 2 (about 10 million files). All the steps to look up the events in the EventIndex, submit the PanDA event picking jobs, monitor them and retry them when timing out (because of long tape staging delays) were executed manually.
The main goal of the Event Picking Service is to perform all these actions automatically. A user would only have to supply to this service all relevant information to find the requested events such as a file containing the run and event numbers, data format of requested data, project name, trigger stream in case raw data are requested and the version of the requested events if other than raw data are in the request.

T. Zh. Bezhanyan

Development of algorithms and web services for the automation of behavioral test data analysis

The report presents the status of the work carried out within the joint project of MLIT and LRB JINR, aimed at developing a web service module for radiation biology tasks, enabling the analysis of behavioral tests performed on small laboratory animals. The module is intended to classify the movement trajectories of laboratory animals, known as search strategies, in the “Morris Water Maze” setup, which is used to evaluate the spatial memory of experimental animals during behavioral testing.
To apply the neural network approach to the classification task, an annotated dataset, including installation files marking and trajectory construction, should be prepared. Video analysis algorithms based on computer vision have been elaborated for this purpose. For the convenience and verification of the correctness of the obtained trajectories, a web service has been developed; it allows one to load, analyze the trajectories and create marked up dataset.

The development is based on the ML/DL/HPC ecosystem of the HybriLIT heterogeneous platform.