Add important details about external memory to docs #11113
Conversation
doc/tutorials/external_memory.rst
Outdated
| (unless specified by the ``extmem_single_page``) . Instead, it caches all batches in the | ||
| external memory and fetch them on-demand. Go to the end of the document to see a | ||
| comparison between :py:class:`~xgboost.QuantileDMatrix` and the external memory version of | ||
| :py:class:`~xgboost.ExtMemQuantileDMatrix`. | ||
| external memory to disk (in a compressed format) and fetches them on-demand. Go to the |
There was a problem hiding this comment.
We use external memory to refer to disk. For the CPU, the "main memory" is the memory we usually talk about, and external memory can refer to anything else from local disk to network storage. This is not an XGBoost convention; it's used elsewhere as well. Computation using "main memory" data is sometimes called "in-core" computing.
Only a disk is supported as an external memory for XGBoost. As a result, from the perspective of a CPU-based algorithm, "external memory" and "disk" are the same thing.
For GPU, the main memory is the device memory, whereas CPU memory and disk are "external" to GPU.
There was a problem hiding this comment.
Thanks. I'm not very familiar with libraries offering larger-than-memory compute, but this is the first time I'm hearing of from-disk-caches as "external memory". I do see there's a whole wikipedia article about it though. "out-of-core" I guess is quite common and typically understood to mean from-disk though. Perhaps it could have that explanation in the .rst file too.
There was a problem hiding this comment.
Maybe a brief intro with a few sentences. But for this PR, could you please help revise or revert the change?
There was a problem hiding this comment.
Undid the changes to the .rst file.
Co-authored-by: Jiaming Yuan <[email protected]>
Not 100% sure that I'm getting all the details right, so @trivialfis please take a careful look.