[
https://issues.apache.org/jira/browse/CASSANDRA-6271?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13833644#comment-13833644
]
Benedict commented on CASSANDRA-6271:
-------------------------------------
bq. Also CFS.removeDeletedColumnsOnly()/removeDroppedColumns(), that are called
during memtable flush
Right, these are the two I spotted that I was worried about, which is why I was
leaning towards just blotting out the Column in the data structure to avoid any
expensive array allocations. Although I hadn't absolutely ruled out any other
places yet. If this is the only spot, moving to a single pass iteration of the
columns for flushing does seem to be a better idea though.
> Replace SnapTree in AtomicSortedColumns
> ---------------------------------------
>
> Key: CASSANDRA-6271
> URL: https://issues.apache.org/jira/browse/CASSANDRA-6271
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Benedict
> Assignee: Benedict
> Labels: performance
> Attachments: oprate.svg
>
>
> On the write path a huge percentage of time is spent in GC (>50% in my tests,
> if accounting for slow down due to parallel marking). SnapTrees are both GC
> unfriendly due to their structure and also very expensive to keep around -
> each column name in AtomicSortedColumns uses > 100 bytes on average
> (excluding the actual ByteBuffer).
> I suggest using a sorted array; changes are supplied at-once, as opposed to
> one at a time, and if < 10% of the keys in the array change (and data equal
> to < 10% of the size of the key array) we simply overlay a new array of
> changes only over the top. Otherwise we rewrite the array. This method should
> ensure much less GC overhead, and also save approximately 80% of the current
> memory overhead.
> TreeMap is similarly difficult object for the GC, and a related task might be
> to remove it where not strictly necessary, even though we don't keep them
> hanging around for long. TreeMapBackedSortedColumns, for instance, seems to
> be used in a lot of places where we could simply sort the columns.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
