Re: Poor performance on windows with logging sinks::file_collector when two file logs use same directory. Thoughts?

From: Niall Douglas (s_sourceforge_at_[hidden])
Date: 2021-06-24 13:26:50

On 23/06/2021 18:39, Nicholas Neumann via Boost wrote:

> I've got two different rotating file logs that point to the same directory
> (with different file name patterns). Is this in general just a bad idea?
> They end up sharing the same file_collector, which seems wrong, so perhaps
> that is a clue that I shouldn't have my logs set up like this.
> 
> In production I've got a service that compresses, archives, and manages the
> size of the logs. But in dev, I don't, so the number of files in the
> directory slowly grew. But the startup time for my program grew much
> faster. On windows the scan_for_files function in the collector has a loop
> that is O(mn), where m is the number of files in the directory, and n is
> the number that matched in previous calls to the scan_for_files function
> 
> This means the scan_for_files for the first rotating file log in the
> directory has no issue (n is 0), but the second can be problematic. It
> iterates over the files in the directory and for each file in the
> directory, it calls filesystem::equivalent on all of the matches from
> previous scan_for_files calls. On windows, filesystem::equivalent is
> particularly heavy, opening handles to both files.
> 
> Thoughts? Is the two file logs getting the same collector the real issue?
> Or is it my pointing two file logs to the same directory? I see some ways
> to mitigate the slowdown in scan_for_files - e.g., filesystem::equivalent
> could be called after all of the method/match_pattern check, but the two
> file logs sharing the same collector feels like the real issue.

The Windows filesystem is much slower than the filesystems on POSIX,
about 70x slower for non-i/o operations. On top of that, the Win32 API
adds considerable overhead above the kernel, and on top of that again,
filesystem abstractions such as C++ Filesystem add a lot more again. It
would not be surprising if naive C++ were 200x-500x slower on Windows
for some filesystem traversal operations than on Linux.

No matter what anyone (apart from the Windows kernel team) does will
Windows filesystem ever be performance competitive with any POSIX
implementation. However, you can close the gap very significantly if you
employ a bare metal abstraction such as LLFIO (disclaimer: it's mine,
but parts are up before WG21 for standardisation). In my work where
there is a Petabyte scale custom DB which is LLFIO based, Windows is
between 20-40% slower than Linux on the same hardware, which is pretty
good given it's a really stupid DB which relies heavily on the
filesystem to be fast.

All modern filesystems perform well with LLFIO with > 10 million files
in the same directory. We somewhat cheat on Windows because we avoid
per-file handle open-close for metadata ops (e.g.
filesystem::equivalent) by using a single syscall kernel globbed
directory enumeration of the parent directory instead, this yields
similar performance to POSIX.

Anyway I suspect all this is overkill for your use case, but I mention
it anyway. Bad filesystem performance on Windows is very fixable, with
some refactoring.

Niall