cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Benedict (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-6694) Slightly More Off-Heap Memtables
Date Tue, 08 Apr 2014 08:15:18 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-6694?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13962691#comment-13962691
] 

Benedict commented on CASSANDRA-6694:
-------------------------------------

bq. I think Jonathan Ellis mentioned that it might be better to reduce usage of the column
names instead of merging cell with column name (if I remember correctly)

I don't recall this suggestion. Perhaps you are referring to the suggestion that we not extract
the cell names from the cell as often as we do, for the purpose of comparison, in order to
reduce garbage production?

bq. Regarding placeholders idea, if we allocate contiguous region for the whole cell we can
just have memory object + 1 int (or was it even short?...) field which marks the end of the
column name at that buffer, as column timestamp is a fixed size long we know exactly where
column value ends, that also helps with spatial locality in most of the 

In this case, this suggestion has much more complex problems:

# More (multiple implementation) virtual method invocations (as shown by CASSANDRA-6993 this
can have meaningfully negative performance implications)
# Major refactor of AbstractType hierarchy to prevent bytebuffer allocation on comparison
# More object allocation in the request threads due to having to re-pack all of any parameters
into a Cell with a single buffer, as opposed to just dropping them in place
# At which point it would make most sense to refactor (and mostly eliminate) the entirety
of CASSANDRA-5417, as we're almost always pumping the result straight into a Cell anyway,
so extracting the components into separate buffers and repacking them into a single buffer
in the Cell is very wasteful

That said, it is *viable*. It has some advantages too: the comparisons between Native and
Buffer cells are much more easily optimised. Many of these changes may well need to happen
in the natural course of things anyway as we optimise the native implementation. But it has
comparatively wide-ranging implications for the current on-heap use case that might be a bit
too much to bite off right now.

bq. if it's not essential then we can do it at the very last stage once we done with all more
important changes which are plenty

I disagree. It makes the patch more complicated to *not* move it around. Because something
is not essential does not mean it is not the better option

> Slightly More Off-Heap Memtables
> --------------------------------
>
>                 Key: CASSANDRA-6694
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-6694
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Benedict
>            Assignee: Benedict
>              Labels: performance
>             Fix For: 2.1 beta2
>
>
> The Off Heap memtables introduced in CASSANDRA-6689 don't go far enough, as the on-heap
overhead is still very large. It should not be tremendously difficult to extend these changes
so that we allocate entire Cells off-heap, instead of multiple BBs per Cell (with all their
associated overhead).
> The goal (if possible) is to reach an overhead of 16-bytes per Cell (plus 4-6 bytes per
cell on average for the btree overhead, for a total overhead of around 20-22 bytes). This
translates to 8-byte object overhead, 4-byte address (we will do alignment tricks like the
VM to allow us to address a reasonably large memory space, although this trick is unlikely
to last us forever, at which point we will have to bite the bullet and accept a 24-byte per
cell overhead), and 4-byte object reference for maintaining our internal list of allocations,
which is unfortunately necessary since we cannot safely (and cheaply) walk the object graph
we allocate otherwise, which is necessary for (allocation-) compaction and pointer rewriting.
> The ugliest thing here is going to be implementing the various CellName instances so
that they may be backed by native memory OR heap memory.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message