A Bytedance engineer identified a significant bottleneck in the Linux kernel's IOmap framework affecting small direct I/O performance on PCIe Gen5 NVMe SSDs. The fix introduces a simplified direct I/O path for small I/O requests (size <= inode blocksize) on EXT4 and XFS, bypassing heavy IOmap overhead. Benchmarks show an improvement from 1.92M to 2.19M IOPS for 4K random reads, with up to 10% gains when using io_uring with high queue depths. The patch has been merged into the VFS subsystem's vfs-7.3.iomap branch and is expected to land in Linux 7.3.
1.7K Impressions