ChrisHegarty commented on issue #13551: URL: https://github.com/apache/lucene/issues/13551#issuecomment-2470361869
Late to the party!!! I want to dig a little further on the distinction between the implementation of preload and prefetch, at least with the mmap implementation. The former touches every page (effectively forcing into the cache), while the latter simply advises of the access pattern. For sequential access, a combination of sequential advice and prefetch works really well, but not so well for random access. Random access, at least for the HNSW graph and the related vector data (be it quantised or not), really needs to be resident in memory. Preload gives us that, at least at the time of opening. Prefect does not. For optimized vector search use cases, assuming enough RAM, I was thinking that a separate `IndexInput::load` could be useful. This would be similar to preload (touches every page), but can be called after opening. It may or may not be super helpful, it mostly depends on how things get setup, etc, and whether or not preload is good enough for your use case. For optimized vector search use cases, assuming enough RAM, I was thinking that a separate `IndexInput::lock` could be useful. Basically a wrapper around `mlock`. The idea being that Vector search perf sucks when there is not enough RAM to keep the graph and vector data resident in memory, so allow to load and lock it, then leave the tradeoff to the user to decide if they want this "fail fast" rather then "search slowly" behaviour (at least for pure vector search). -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
