Re: speeding up installs
On Fri, Jun 07, 2019 at 10:52:43AM -0700, Russ Allbery wrote:
> Adam Borowski <kilobyte@angband.pl> writes:
>
> > There are two massive recent improvements:
> > * eatmydata helps a lot, and eating a power-lossed install is not a loss
>
> We did an experiment at work about four years ago with using eatmydata
> inside a D-I run, and had to revert because we ran into no end of problems
> with zero-length files after the first system reboot, even though we did
> an explicit sync before reboot.
>
> This is potentially obsolete data (this was with Ubuntu 12.04, IIRC), but
> it's made me fairly cautious of eatmydata ever since.
>
> Also, I think your proposal is not to use LD_PRELOAD and other hacks but
> to instead have a native package installation mode that understands when
> to sync, and perhaps that wouldn't have the same behavior.
Yeah, I use archive_write_disk (going to drop it probably, though), and
ordinary write(). Not even a single mmap() to write or anything fancy --
libarchive doesn't write into user-provided buffers so dax shenanigans or
such won't get us anything; can just let the kernel write() without
bothering to optimize.
If a filesystem can screw sync() after a bunch of plain open() write()
close()s, there's nothing you can sanely do.
Meow!
--
⢀⣴⠾⠻⢶⣦⠀
⣾⠁⢠⠒⠀⣿⡁ Sometimes you benefit from delegating stuff. For example,
⢿⡄⠘⠷⠚⠋⠀ this way I get to be a vegetarian.
⠈⠳⣄⠀⠀⠀⠀
Reply to: