Lore framework is used for dumping input data of a gpu executor to disk so that it can be replayed in local environment for performance analysis.
Lore framework is used for dumping input data of a gpu executor to disk so that it can be
replayed in local environment for performance analysis.
When RapidsConf.TAG_LORE_ID_ENABLED is set, during the planning phase we will tag a lore
id to each gpu operator. Lore id is guaranteed to be unique within a query, and it's supposed
to be same for operators with same plan.
When RapidsConf.LORE_DUMP_IDS is set, during the execution phase we will dump the input
data of gpu operators with lore id to disk. The dumped data can be replayed in local
environment. The dumped data will reside in RapidsConf.LORE_DUMP_PATH. For more details,
please refer to docs/dev/lore.md.