MMAP makes machine swap a lot

[Scooletz]

On Windows, memory mapped file occasionally makes a computer swap a lot and makes Paprika behave terribly ppor. With a lot of memory committed, when the size of the mmaped region breaches the space on system disk, it starts to stall. Investigate this behavior and provide a way to fix it. This happens in a scenario of an import where mmaped file is used without any write or flushes and FlushViewOfAFile is called only at the end to be followed by FlushFileBuffers. Still, according to my understanding system should flush what is needed on it's own so that the swap is not affected.

[ssonthal]

The first solution that can help here is related to dynamically allocating memory mapped files using NtCreateSection APIs. This will reduce the size of the memory mapped files to be written to disk and saving a lot of RAM.

Secondly, during imports, we can try writing calling FlushViewOfFile and FlushFileBuffers after every 10 MB (size-wise) or 5 seconds (duration-wise) instead of doing it all at the end.

Let me know what you think.

[Scooletz]

This could be issue that happens only when there's no Flush. We could test it and disable the mmap option and go with file writes and flush at the end. This requires a repro though and then more testing.

[Scooletz]

@damian-orzechowski reports that occasional .FlushForce (MSYNC + FSYNC) decreases the overhead a lot. The call itself though is time consuming. One take could be to run in a background thread when running the Snap Sync and/or Importer

[ssonthal]

This could be issue that happens only when there's no Flush. We could test it and disable the mmap option and go with file writes and flush at the end. This requires a repro though and then more testing.

@Scooletz Can we try this? You can create an issue or a lot of sub-issues for this maybe. And I will try to work part-by-part to remove dependency on mmap and use fsync, read, write instead.