[TAJO-584] Improve distributed merge sort

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.8.0
• Component/s: distributed query plan, Physical Operator
• Assignee: Hyunsik Choi
• Reporter: Hyunsik Choi
• Created: 04/Feb/14 14:04
• Updated: 10/Feb/14 16:54
• Resolved: 10/Feb/14 16:49

Description

In Tajo, sort operator is similar to merge sort, and it works in a distributed manner. The first sort phase sorts each fragment in local machine, the intermediate data are shuffled in range partition, and then the second sort phase in each node sorts the range-partitioned data.

However, the second sort phase reads all shuffled data via one scanner. It misses the opportunity to exploit already-sorted data. This patch improves the second sort phase to merge directly multiple already-sorted intermediate data sets. It significantly reduces the response time of sort queries.

I carried out some simple benchmark with the following query on TPC-H 100GB data sets:

select l_orderkey from lineitem order by l_orderkey;

The lineitem table occupies 75GB. The query response time are dramatically reduced from 480 to 260 secs. This patch exploits the design of TAJO-36. So, this patch requires TAJO-36.

Issue Links

• depends upon: TAJO-36 Improve ExternalSortExec with N-merge sort and final pass omission