IO_uring Async Discard Submitted For Linux 6.12
Jens Axboe submitted the block and IO_uring changes already for the now-open Linux 6.12 merge window. Most notable from this Linux I/O work is adding async discard support to IO_uring.
IO_uring async discard support was sent in as one of the feature pulls aiming for Linux 6.12. Applications can now issue async discards rather than blocking sync ioctl discards. On slower storage devices this async discard support can make a big impact on performance with cases going from like 14K IOPS with sync discard to around 56K IOPS with async discard. Or some really wild gains in synthetic null block device testing:
More details for those interested within the IO_uring async discard pull.
Separately was the main IO_uring pull for Linux 6.12 with support for absolute timeouts, efficiency improvements when dealing with huge pages, support for a minwait mode, support for incremental buffer consumption, support for cloning registered buffers from one ring to another, and other improvements to the awesome IO_uring functionality.
The block pull request for Linux 6.12 includes RAID5 MD performance optimizations, metadata support validation for NVMe, asynchronous namespace scanning for NVMe, and a new maintainer for the BFQ scheduler code that had been unmaintained for a while.
IO_uring async discard support was sent in as one of the feature pulls aiming for Linux 6.12. Applications can now issue async discards rather than blocking sync ioctl discards. On slower storage devices this async discard support can make a big impact on performance with cases going from like 14K IOPS with sync discard to around 56K IOPS with async discard. Or some really wild gains in synthetic null block device testing:
More details for those interested within the IO_uring async discard pull.
Separately was the main IO_uring pull for Linux 6.12 with support for absolute timeouts, efficiency improvements when dealing with huge pages, support for a minwait mode, support for incremental buffer consumption, support for cloning registered buffers from one ring to another, and other improvements to the awesome IO_uring functionality.
The block pull request for Linux 6.12 includes RAID5 MD performance optimizations, metadata support validation for NVMe, asynchronous namespace scanning for NVMe, and a new maintainer for the BFQ scheduler code that had been unmaintained for a while.
1 Comment