Arvid Norberg

Ok. When memory mapping a file, that file itself is backing the pages, so (presumably) they won't affect the pagefile, or be relevant for purposes of extending the pagefile. That...

My impression is that Linux is one of the most sophisticated kernels when it comes to managing the page cache. Windows is having more problems where the system won't flush...

`MADV_DONTNEED` will destroy the contents of dirty pages, so I don't think that's an option, but there's `MADV_COLD` in newer versions of linux.

> For the mmap issue, how about creating many smaller maps instead of the large map of the entire file I don't think it's obvious that it would make a...

a quick summary is that libtorrent 2.0 uses memory mapped files. On all modern operating systems, other than windows, use a unified page cache, at the block device level, where...

also, when reporting problems in libtorrent (specifically the mmap-disk backend), *just* pointing to vmstats numbers suggesting the kernel decided to use a lot of physical memory for disk cache is...

to be clear, it's the kernel that decides what the physical memory is best used for. Do you have a reason to believe the kernel page cache performed worse than...

> The docs for libtorrent's cache_size say the cache is shared between socket buffers and file cache, so just because it's full doesn't mean it's full of file data (@arvidn...

> An counter argument could be kernel is dumb on telling which page more important when memory pressure comes I wasn't making an argument, I was asking a question. Are...

> Afaik, the default Linux page cache (not those advanced cache such as zfs arc) doesn't track per page hit/miss ratio or if has been accessed recently, You could try...