Cephalocon 2022

In a distributed storage environment like Ceph, client operations (foreground) and background operations such as recovery, scrubs, snap trim etc., compete for resources and depending on the situation, one can overwhelm the other without any QoS or other throttling mechanisms in place. Ceph's use of mClock to achieve QoS is based on an implementation of the dmClock algorithm. Recent refinements made to the mClock scheduler implementation in the Ceph code base leverages the full potential of the algorithm. This talk presents the results of a study that compared the existing 'weighted priority queue' scheduler with the 'mclock scheduler' and the refinements that were made to the mClock scheduler in Ceph. The study performed tests with client operations and background operations like recoveries and scrubs. The results show that mClock scheduler is able to provide better QoS to the different classes of operations when compared to the 'weighted priority queue'. The talk also introduces config profiles that are classified based on whether preference is to be given to client operations, background operations or provide a balance between all the operations. When applied, the profile automatically sets the QoS parameters in the mClock server thus reducing complexity from users.