Uploaded image for project: 'Phoenix'
  1. Phoenix
  2. PHOENIX-2126

Improving performance of merge sort by multi-threaded and minheap implementation

    XMLWordPrintableJSON

    Details

    • Type: Improvement
    • Status: Patch Available
    • Priority: Major
    • Resolution: Unresolved
    • Affects Version/s: 4.1.0, 4.2.0
    • Fix Version/s: None
    • Component/s: None
    • Labels:
      None

      Description

      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';
      

      Query to benchmark:-

      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;
      

      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)

        Attachments

        1. PHOENIX-2126_v3.patch
          33 kB
          Ankit Singhal
        2. PHOENIX-2126_v2.0.patch
          32 kB
          Ankit Singhal
        3. PHOENIX-2126_v1.0.patch
          28 kB
          Ankit Singhal

          Activity

            People

            • Assignee:
              ankit.singhal Ankit Singhal
              Reporter:
              ankit.singhal Ankit Singhal
            • Votes:
              2 Vote for this issue
              Watchers:
              10 Start watching this issue

              Dates

              • Created:
                Updated: