Speaker
Andreas Joachim Peters
(CERN)
Description
In preparation for Run-3 we have faced the following problem: we have to balance the usage of IO resources between individual activities, which has led to the implementation of IO priorities and bandwidth regulation policies. While commissioning the ALICEO2 EOS instance we have observed, that write performance using the buffer cache is a bottleneck on storage nodes. Direct IO helps to improve write performance on storage nodes and additionally to reduce tails in data upload times by experiment DAQ systems. The presentation will explain and implementation and how to use these features in production.
Primary author
Andreas Joachim Peters
(CERN)
