Hans Zeller created TRAFODION-2437:
--------------------------------------
Summary: Implement a monitor for communications between ESPs of a
fragment
Key: TRAFODION-2437
URL: https://issues.apache.org/jira/browse/TRAFODION-2437
Project: Apache Trafodion
Issue Type: New Feature
Components: sql-exe
Reporter: Hans Zeller
Assignee: Hans Zeller
Priority: Minor
Today, the ESPs that are executing parallel fragment instances can't talk to
each other. They all do their work independently and may finish at different
times. This works fine for normal parallel algorithms, but it somewhat
restricts what we can do in parallel. Sometimes, a communication or maybe just
a simple synchronization would be desirable. Examples:
- Make all ESPs wait until the last ESP has finished before proceeding. This
could create a global blocking operator, for example when we need a blocking
operator to prevent the Halloween problem.
- Dynamic range-repartitioning: After doing local sorts, negotiate common split
boundaries to perform a range-repartitioning, so that we can produce a globally
sorted result in parallel.
- Communication between parallel instances of TMUDFs. Some TMUDFs may need
similar communication, which could allow them to do more things in parallel.
When implementing this, we may also need to pay some attention to potential
deadlocks, although with blocking operators like sorts there should not be any
deadlock issues.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
