[EXT] Re: [PATCHv11 0/9] write hints with nvme fdp and scsi streams

[Keith Busch]

On Thu, Nov 14, 2024 at 07:07:10AM +0100, Christoph Hellwig wrote:
> On Thu, Nov 14, 2024 at 10:51:09AM +1100, Dave Chinner wrote:
> > SGI wrote an entirely new allocator for XFS whose only purpose in
> > life is to automatically separate individual streams of user data
> > into physically separate regions of LBA space.
> 
> One of the interesting quirks of streams/FDP is that they they operate
> on (semi-)physical data placement that is completely decoupled from LBA
> space.  

That's not particularly interesting about FDP. All conventional NAND
SSDs operates that way. FDP just reports more stuff because that's what
people kept asking for. But it doesn't require you fundamentally change
how you acces it. You've singled out FDP to force a sequential write
requirement that that requires unique support from every filesystem
despite the feature not needing that.

We have demonstrated 40% reduction in write amplifcation from doing the
most simplist possible thing that doesn't require any filesystem or
kernel-user ABI changes at all. You might think that's not significant
enough to let people realize those gains without more invasive block
stack changes, but you may not buying NAND in bulk if that's the case.