[
https://issues.apache.org/jira/browse/PHOENIX-2126?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14975439#comment-14975439
]
James Taylor commented on PHOENIX-2126:
---------------------------------------
Thanks for the patch, [~ankit.singhal]. There's some really excellent work
there - we really appreciate it. It'd be good to break up the patch into more
manageable pieces as it's pretty sizable. I think different parts of your patch
may be applicable for different query optimizations. In general, though, we
don't want to pull over all rows from the server into a memory mapped file on
the client for all the cases doing merge sort. In the case of a multi billion
row table, just imagine the disk space that would be required (FWIW, we used to
do this when a SELECT * query was done with our spooling, but changed it for
just that reason).
Here's an idea on how to get this committed in pieces:
- Use a priority queue in MergeSortResultIterator. This will greatly reduce the
number of key comparisons we're doing. The queue depth would be bounded by the
number of child iterators, so memory consumption would not be an issue.
- Use your parallelized two level heap implementation to perform a client-side
sort when it's necessary. This would be a big performance boost in these
situations as we're doing this client side sort single threaded currently (and
we're already buffering on the client side so there's no additional cost for
that).
Thoughts?
> Improving performance of merge sort by multi-threaded and minheap
> implementation
> --------------------------------------------------------------------------------
>
> Key: PHOENIX-2126
> URL: https://issues.apache.org/jira/browse/PHOENIX-2126
> Project: Phoenix
> Issue Type: Improvement
> Affects Versions: 4.1.0, 4.2.0
> Reporter: Ankit Singhal
> Attachments: PHOENIX-2126_v1.0.patch, PHOENIX-2126_v2.0.patch,
> PHOENIX-2126_v3.patch
>
>
> {code}
> CREATE TABLE IF NOT EXISTS test (
> dim1 INTEGER NOT NULL,
> A.B INTEGER,
> A.M DECIMAL,
> CONSTRAINT PK PRIMARY KEY
> (dim1))
> SALT_BUCKETS =256,DEFAULT_COLUMN_FAMILY='A';
> {code}
> *Query to benchmark:-*
> {code}
> select dim1,sum(b),sum(m) from test where Datemth>=201505 and Datemth<=201505
> and dim1 IS NOT NULL group by dim1 order by sum(m) desc nulls last limit 10;
> {code}
> *current scenario:-*
> *CASE 1: * consider the case when dim1 is high cardinality attribute (10K+)
> and table have salt bucket set to 256, we will get 256 iterators from above
> query at the client and MergeSortRowKeyResultIterator has to merge these 256
> iterators with single thread. So let's say each iterator has 10k tuples
> returned, then merge sort needs to merge 2.5M tuples which will be costly if
> it is done with single thread and the query spend most of its time on client
> *CASE 2: * consider the case when dim1 is high cardinality attribute (10K+)
> and table have salt bucket set to 1, we will get 1 iterator from above query
> at the client and MergeSortRowKeyResultIterator doesn't need to merge
> anything. Here, it is fine with single thread.
> *CASE 3: * consider the case when dim1 is low cardinality attribute (10-100)
> and table have salt bucket set to 256, we will get 256 iterator from above
> query at the client and MergeSortRowKeyResultIterator has to merge these 256
> iterators with single thread. here the single thread is also fine as he has
> to merge only 2560 tuples.
> *Solution for case1 problem is:-*
> Optimized the implementation of merging 'n'-sorted iterators(having 'm'
> tuples) by using "min heap" which optimizes the time complexity from
> O(n2m) to O(nmLogn) (as heapify takes (Logn) time).
> And, By using multiple-threads('t') to process group of iterators which
> further optimized the complexity to
> T(nm)=T(nm)/t+T(t)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
