...

As RfK traverses the Katana scene graph it will process locations in parallel and pass them to RenderMan for rendering. By default each new child location retrieved from Katana will spawn a new thread for processing. This default behavior is unlikely to give the optimal performance for any given scene, however there are tools available PrmanObjectStatements includes settings for tuning the efficiency and the level of parallelism for a particular pipeline (e.g. your studio's standard Op Chain) or various types of scenes (e.g. many instances, many textures, many primitives, or perhaps all three).

...

By default siblings of a Katana scenegraph location will be evicted from a thread's runtime memory when a location is processed for conversion to the Rix scenegraph. Setting this flag false prevents sibling eviction at this location in the event it is known that those siblings are best kept in the cache for subsequent access.

The last tool is the flattenInstanceSource setting.