Re: [PATCH 2/2] core.fsyncobjectfiles: batch disk flushes

On Tue, Aug 24, 2021 at 10:38 PM Christoph Hellwig [off-list ref] wrote:

On Wed, Aug 25, 2021 at 01:51:32AM +0000, Neeraj Singh via GitGitGadget wrote:

quoted

From: Neeraj Singh <redacted>

When adding many objects to a repo with core.fsyncObjectFiles set to
true, the cost of fsync'ing each object file can become prohibitive.

One major source of the cost of fsync is the implied flush of the
hardware writeback cache within the disk drive. Fortunately, Windows,
MacOS, and Linux each offer mechanisms to write data from the filesystem
page cache without initiating a hardware flush.

This patch introduces a new 'core.fsyncObjectFiles = 2' option that
takes advantage of the bulk-checkin infrastructure to batch up hardware
flushes.

Another interesting way to flush on linux would be the syncfs call,
which syncs all files on a file system.  Once you write more than
handful or two of files that tends to win out over a batch of fsync
calls.

I'd expect syncfs to suffer from the noisy-neighbor problem that Linus
alluded to on the big
thread you kicked off.  The equivalent call on Windows currently
requires administrative
privileges (I'm really not sure exactly why, perhaps we should change that).

If someone adds a more targeted bulk sync interface to the Linux
kernel, I'm sure Git could be
changed to use it. Maybe an fcntl(2) interface that initiates
writeback and registers completion with an
eventfd.