Speaker
Description
This contribution introduces the job optimizer service for the next-
generation ALICE grid middleware, JAliEn (Java Alice Environment).
It is a continuous service running on central machines and is essentially
responsible for splitting jobs into subjobs, to then be distributed and
executed on the ALICE grid. There are several ways of creating subjobs
based on various strategies relevant to the aim of any particular grid job.
Therefore a user has to explicitly declare that a job is to be split, and
also define the strategy to be used. The new job optimizer service aims to
retain the old ALICE grid middleware functionalities from the user’s point
of view while increasing the performance and throughput. One aspect
of increasing performance is looking at how the job optimizer interacts
with the job queue database. A different way of describing subjobs in the
database is presented, to minimize resource usage. There is also a focus on
limiting communications with the database, as this is already a congested
area. Furthermore, a new solution to splitting based on the locality of job
input data will be presented, aiming to split into subjobs more efficiently,
therefore making better use of resources on the grid to further increase
throughput. Added options for the user regarding splitting by locality,
such as setting a minimum limit for a subjob size, will also be explored.
| Consider for long presentation | No |
|---|
