Thread (9 messages) read the whole thread 9 messages, 5 authors, 2025-02-13

Re: [PATCH 3/3] apple-nvme: defer cache flushes by a specified amount

From: Christoph Hellwig <hch@lst.de>
Date: 2025-02-13 06:20:18
Also in: asahi, linux-nvme, lkml

On Tue, Feb 11, 2025 at 01:25:59PM -0500, Alyssa Rosenzweig wrote:
From: Jens Axboe <axboe@kernel.dk>

Cache flushes on the M1 nvme are really slow, taking 17-18 msec to
complete. This can slow down workloads considerably, pure random writes
end up being bound by the flush latency and hence run at 55-60 IOPS.

Add a deferred flush work around to provide better performance, at a
minimal risk. By default, flushes are delayed at most 1 second, but this
is configurable.

With this work-around, a pure random write workload runs at ~12K IOPS
rather than 56 IOPS.
Just as last time this really is not a driver feature.  Cache flushes
are slow on consumer hardware, it's just apple is worse than usual.
Breaking file system transactional guarantee by ignoring data integrity
command in the driver is a no-go.

If we want to allow an opt-in policy for those whole feel adventurous,
it belongs into the core flush state machine.  Fortunately the patch
author seems qualified to touch that :)


Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help