[IMPALA-9378] CPU usage for runtime profiles with multithreading

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: Backend
• Labels: multithreading performance
• Epic Link: Multithreading upgrade path for large clusters
• Epic Color: ghx-label-13
• Assignee: Unassigned
• Reporter: Tim Armstrong
• Created: 13/Feb/20 15:56
• Updated: 09/Feb/21 04:45

Description

David Rorke reports that significant amounts of time can be spent on the runtime profile with higher values of mt_dop. This can impact query performance from the client's point of view since profile serialisation is on the critical path for closing the query. Also serialising the profile for the webserver holds the ClientRequestState's lock, so can block query progress.

We should figure out how to make this more efficient.

Issue Links

• is duplicated by: IMPALA-4603 Serializing large query profiles as part of ImpalaServer::UnregisterQuery can potentially cause slow downs

Sub-Tasks

• Serialize query profile asynchronously: Resolved
• Lazily convert and/or cache different representations of the query profile: Resolved
• Optimise RuntimeProfile::ToThrift(): Resolved
• Move instance profile operations to executors: Resolved
• Prototype denser runtime profile implementation: Resolved
• Incrementally compute averaged profile: Resolved
• Switch to aggregated runtime profile representation: Open
• JSON profiles are mostly space characters: Resolved
• Utility to pretty-print thrift profiles at various levels: Resolved