[
https://issues.apache.org/jira/browse/HBASE-9969?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13825816#comment-13825816
]
Matt Corgan commented on HBASE-9969:
------------------------------------
Regarding adding 10 cols/row:
{quote}I tried this on my laptop but seems your case above is even faster than
before. Maybe there is something wrong with my environment. I will try it on my
devbox tomorrow.{quote}yes, you are right. i was profiling this weekend and
confirmed the current heap is handling that situation favorably. still good to
test to make sure we don't lose this aspect!
I made a stripped down version of the PriorityQueue based heap to compare with
the LoserTree. It adds some counters to track the number of KV comparisons
which is interesting to see. I was seeing that PQ is faster for next(),
especially with just 1 file, and LT is faster for reseek(). I'll try to post a
patch tonight.
I was paying particular attention to this code at KeyValueHeap:103
{code}
KeyValueScanner topScanner = this.heap.peek();
if (topScanner == null ||
this.comparator.compare(kvNext, topScanner.peek()) >= 0) {
this.heap.add(this.current);
this.current = pollRealKV();
}
{code}
I can't figure out why we need to do a heap.add() and pollRealKV when
topScanner==null. I actually removed the topScanner==null check from the above
and the single file scanner was 50% faster. The whole test suite passed, so
either it's not necessary, or we could use another unit test. Maybe it has
something to do with LazySeek?
> Improve KeyValueHeap using loser tree
> -------------------------------------
>
> Key: HBASE-9969
> URL: https://issues.apache.org/jira/browse/HBASE-9969
> Project: HBase
> Issue Type: Improvement
> Components: Performance, regionserver
> Reporter: Chao Shi
> Assignee: Chao Shi
> Fix For: 0.98.0, 0.96.1, 0.94.15
>
> Attachments: 9969-0.94.txt, hbase-9969-v2.patch, hbase-9969-v3.patch,
> hbase-9969.patch, hbase-9969.patch, kvheap-benchmark.png, kvheap-benchmark.txt
>
>
> LoserTree is the better data structure than binary heap. It saves half of the
> comparisons on each next(), though the time complexity is on O(logN).
> Currently A scan or get will go through two KeyValueHeaps, one is merging KVs
> read from multiple HFiles in a single store, the other is merging results
> from multiple stores. This patch should improve the both cases whenever CPU
> is the bottleneck (e.g. scan with filter over cached blocks, HBASE-9811).
> All of the optimization work is done in KeyValueHeap and does not change its
> public interfaces. The new code looks more cleaner and simpler to understand.
--
This message was sent by Atlassian JIRA
(v6.1#6144)
