[
https://issues.apache.org/jira/browse/MESOS-3112?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14635182#comment-14635182
]
Jan Schlicht commented on MESOS-3112:
-------------------------------------
The approach is to store the last access time in FetcherProcess::Cache::Entry
instances. Every time a cache entry is accessed, this value will be set to
Clock::now().
FetcherProcess::Cache::selectVictims(const Bytes& requiredSpace) can then
implement a MRU-retention by creating a sorted list (a std::set which a custom
Compare functor) of all cache entries and select the LRUs by traversing this
list until requiredSpace could be freed.
> Fetcher should perform cache eviction based on cache file usage patterns.
> -------------------------------------------------------------------------
>
> Key: MESOS-3112
> URL: https://issues.apache.org/jira/browse/MESOS-3112
> Project: Mesos
> Issue Type: Improvement
> Components: fetcher
> Reporter: Bernd Mathiske
> Assignee: Jan Schlicht
> Labels: mesosphere, performance
>
> Currently, the fetcher uses a trivial strategy to select eviction victims: it
> picks the first cache file it finds in linear iteration. This means that
> potentially a file that has just been used gets evicted the next moment. This
> performance loss can be avoided by even the simplest enhancement of the
> selection procedure.
> Proposed approach: determine an effective yet relatively uncomplex and quick
> algorithm and implement it in `FetcherProcess::Cache::selectVictims(const
> Bytes& requiredSpace)`. Suggestion: approximate MRU-retention somehow.
> Unit-test what actually happens!
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
