On Tue, Feb 07, 2023 at 04:01:41PM +0530, Nitesh Shetty wrote: > On Mon, Feb 06, 2023 at 08:49:15PM +0800, Ming Lei wrote: > > On Mon, Feb 06, 2023 at 10:00:20AM +0000, Hans Holmberg wrote: > > > I think we're missing a flexible way of routing random-ish > > > write workloads on to zoned storage devices. Implementing a UBLK > > > target for this would be a great way to provide zoned storage > > > benefits to a range of use cases. Creating UBLK target would > > > enable us experiment and move fast, and when we arrive > > > at a common, reasonably stable, solution we could move this into > > > the kernel. > > > > Yeah, UBLK provides one easy way for fast prototype. > > > > > > > > We do have dm-zoned [3]in the kernel, but it requires a bounce > > > on conventional zones for non-sequential writes, resulting in a write > > > amplification of 2x (which is not optimal for flash). > > > > > > Fully random workloads make little sense to store on ZBDs as a > > > host FTL could not be expected to do better than what conventional block > > > devices do today. Fully sequential writes are also well taken care of > > > by conventional block devices. > > > > > > The interesting stuff is what lies in between those extremes. > > > > > > I would like to discuss how we could use UBLK to implement a > > > common FTL with the right knobs to cater for a wide range of workloads > > > that utilize raw block devices. We had some knobs in the now-dead pblk, > > > a FTL for open channel devices, but I think we could do way better than that. > > > > > > Pblk did not require bouncing writes and had knobs for over-provisioning and > > > workload isolation which could be implemented. We could also add options > > > for different garbage collection policies. In userspace it would also > > > be easy to support default block indirection sizes, reducing logical-physical > > > translation table memory overhead. > > > > > > Use cases for such an FTL includes SSD caching stores such as Apache > > > traffic server [1] and CacheLib[2]. CacheLib's block cache and the apache > > > traffic server storage workloads are *almost* zone block device compatible > > > and would need little translation overhead to perform very well on e.g. > > > ZNS SSDs. > > > > > > There are probably more use cases that would benefit. > > > > > > It would also be a great research vehicle for academia. We've used dm-zap > > > for this [4] purpose the last couple of years, but that is not production-ready > > > and cumbersome to improve and maintain as it is implemented as a out-of-tree > > > device mapper. > > > > Maybe it is one beginning for generic open-source userspace SSD FTL, > > which could be useful for people curious in SSD internal. I have > > google several times for such toolkit to see if it can be ported to > > UBLK easily. SSD simulator isn't great, which isn't disk and can't handle > > real data & workloads. With such project, SSD simulator could be less > > useful, IMO. > > > > > > > > ublk adds a bit of latency overhead, but I think this is acceptable at least > > > until we have a great, proven solution, which could be turned into > > > an in-kernel FTL. > > > > We will keep improving ublk io path, and I am working on ublk > > copy. Once it is done, big chunk IO latency could be reduced a lot. > > > > Just curious, will this also involve running do_splice_direct*() in async style > like normal async read/write, instead of offloading to iowq context ? Follows the idea: - adding new type of buffer(splice buffer) to io_uring, this buffer will be populated into bvec table(reusing io_mapped_ubuf) by passing (splice_fd, offset, len) from SQE. - The buffer is filled from ublk ->read_splice() with help of splice_direct_to_actor() over direct pipe, probably we can add one private splice flag to just allow ublk ->read_splice() to be available in kernel(io_uring) & direct pipe - It requires the pipe buffer ownership not transferred, so nop_pipe_buf_ops is needed for such usage, and this way is pretty fine for ublk & fuse. - The buffer can be allocated & populated from ->prep() of io_uring rw/net, then handled just like READ[WRITE]_FIXED. So it is like normal async read/write, then two pin pages are avoided, and one time of io data copy is saved. This way is also flexible to allow read/write over any part of the buffer. Thanks, Ming
