[
https://issues.apache.org/jira/browse/IMPALA-9433?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Joe McDonnell resolved IMPALA-9433.
-----------------------------------
Fix Version/s: Impala 4.1.0
Resolution: Fixed
> Change FileHandleCache from using a multimap to an unordered_map
> ----------------------------------------------------------------
>
> Key: IMPALA-9433
> URL: https://issues.apache.org/jira/browse/IMPALA-9433
> Project: IMPALA
> Issue Type: Improvement
> Components: Backend
> Affects Versions: Impala 3.4.0
> Reporter: Joe McDonnell
> Assignee: Gergely Fürnstáhl
> Priority: Minor
> Labels: ramp-up
> Fix For: Impala 4.1.0
>
>
> The file handle cache can contain multiple file handles per filename.
> Currently it uses a std::multimap, where the file handles for each filename
> are a contiguous set of entries. A lookup will find the beginning of that
> range and then iterate through it to find a free one.
> A multimap is implemented as a red-black tree with O(log(N)) lookup, so we
> should be able to improve this by using a hashtable-based structure such as
> unordered_map/unordered_multimap with O(1) lookup.
> Another optimization would be to add an intermediary structure for each
> filename and hold all the file handles for that file name in a linked list.
> Lookup would find this intermediary structure by looking up the filename,
> then it would iterate. In the current method, the key/value pair for each
> file handle must store a copy of the filename string as the key, even for
> duplicates. With the intermediary structure, it would store the filename once
> per unique filename.
> It also looks like the LRU list would benefit from being a Boost intrusive
> list ([https://www.boost.org/doc/libs/1_64_0/doc/html/intrusive.html]). Every
> file handle is always in the LRU list, so a std::list has a higher memory
> overhead and requires more memory accesses. It also complicates the code,
> because the FileHandleEntry needs to store a LruListType::iterator to its
> location in the LRU list.
> These optimizations are low priority, but they provide good ramp-up for some
> C++ concepts/APIs.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
