[ 
https://issues.apache.org/jira/browse/IMPALA-5081?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17757684#comment-17757684
 ] 

Yida Wu commented on IMPALA-5081:
---------------------------------

[~MikaelSmith] Several solutions on my mind:
1. we have the optimization level in the key, so that multiple cache entries 
with different optimization levels of the fragment exist. (it may affect the 
hit rate by evicting other entries earlier and leaving some useless entries in 
the cache)
2. we have the optimization level in the cache entry content, replace when hits 
and if the current optimization level is different and better (not sure how to 
define better, or maybe replace with the latest). (would it be costly to update 
the entry?)
3. do nothing in the code. The user tries different optimization levels with 
codegen caching disabled for the given query, and find out the best solution. 
Then restart the server to refill the codegen caching with the best 
optimization level.

> Expose IR optimization level via query option
> ---------------------------------------------
>
>                 Key: IMPALA-5081
>                 URL: https://issues.apache.org/jira/browse/IMPALA-5081
>             Project: IMPALA
>          Issue Type: Improvement
>          Components: Backend
>            Reporter: Michael Ho
>            Assignee: Michael Smith
>            Priority: Minor
>              Labels: codegen
>
> Certain queries may spend a lot of time in the IR optimization. Currently, 
> there is a start-up option to disable optimization in LLVM. However, it may 
> be of inconvenience to users to have to restart the entire Impala cluster to 
> just use that option. This JIRA aims at exploring exposing a query option for 
> users to choose the optimization level for a given query (e.g. we can have a 
> level which just only have a dead code elimination pass or no optimization at 
> all).



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to