I am trying to cache out geometry from a sim via “ROP Geometry Output” but I don't have enough memory to use “All frames in one batch”. If I define a batch size myself, it will cache out the frames I defined, and then stalls, and never cache out the next batch.

Is there a way to tell “ROP Geometry output” that it have to cache from start to end, without needing to hold everything in memory?

Most PDG jobs are run out of process, with one process per work item. Simulations are usually run with All Frames in One Batch because you generally need the simulation to run in a single process for the entire frame range. If you have a 240 frame simulation and you set Frames per Batch to 120, PDG will start two independent processes: one for frames 1 - 120, and a second one for frames 121 to 240. The second process, which has no idea about the first or its results, will have to simulate frames 1 - 120 anyway in order to be able to begin writing out frames 121.

PDG doesn't determine what's kept in memory or discarded during the job's execution - it just cooks the underlying ROP Geometry node, with whatever settings it has and whatever settings are on the SOP network the ROP points to. It's the same as if you put down a Geometry ROP in /out, and clicked the Render button on it. You may need to play around with Caching options on your simulations DOP network.

Oh I see, so “DOP count” * “DOP cache memory” * “Wedge count or CPU core count (What ever is lowest)” should never exceed the max memory that is available?

When I use more memory this way, then is available in the system, Houdini closes down, without any crash message, is that normal behavior?

Can you attach a .hip file that demonstrates what you're trying to do?