[IMPALA-2651] codegen overhead can be high

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: Impala 2.2, Impala 2.3.0
• Component/s: Backend
• Labels: codegen performance
• Target Version: Product Backlog
• Assignee: Michael Smith
• Reporter: Silvius Rus
• Created: 09/Nov/15 06:32
• Updated: 27/Jul/23 21:11
• Resolved: 27/Jul/23 21:11

Description

We received reports of excessive codegen compilation/optimization times for very large expressions generated by visualization tools.

We should:

1. Expose codegen optimization levels as query options. Currently there is only an all or nothing codegen query option. It's likely that overly complex expressions such as hundreds of cascading conditions take very long and benefit very little from an O2 optimization level, but they could still run significantly faster even at O0 or O1 versus interpreted.
2. Consider dropping to O1 (or turn off riskier passes individually) automatically for very large expressions.
3. Consider parameterizing the compilation duration time limits and set a reasonable default, say 10 seconds. Either disable codegen or reduce it to, say, O0 if compilation takes longer than the preset limit.

Workaround
In some cases disabling codegen can help.

SET disable_codegen=true;

Issue Links

• is duplicated by: IMPALA-3262 Investigate Codegen Performance
• relates to: IMPALA-3259 Codegen is not cancellable and can use a lot of CPU and memory
• relates to: IMPALA-5443 Consider automatically disabling codegen per ExecNode based on planner estimates
• relates to: IMPALA-5081 Expose IR optimization level via query option
• relates to: IMPALA-5444 Asynchronous code generation