Hello, This is the second version of my patches to speed up orphan inode handling in ext4. Orphan inode handling in ext4 is a bottleneck for workloads which heavily excercise truncate / unlink of small files as they contend on global s_orphan_mutex (when you have fast enough storage). This patch set implements new way of handling orphan inodes - instead of using a linked list, we store inode numbers of orphaned inodes in a file which is possible to implement in a more scalable manner than linked list manipulations. See description of patch 2/3 for more details. The patch set achieves significant gains both for a micro benchmark stressing orphan inode handling (truncating file byte-by-byte, several threads in parallel) and for reaim new_fserver workload. As a highlight, microbenchmark runtime for 128 threads is reduced from original 160 s down to 71 s, which is also the time it takes the benchmark to run when orphan inode handling is completely disabled. For full numbers you can check commit logs of patches 2/3 and 3/3. You can also check my presentation from Vault at http://events.linuxfoundation.org/sites/events/files/slides/ext4-scaling.pdf for graphs from tests. I'm happy for any review, thoughts, ideas about the patches. The kernel part of the feature is complete, I have also implemented full support in e2fsprogs. That still needs some debugging (especially the e2fsck part) but support in mke2fs or tune2fs is fine. I'll post these as a separate patch so that people can try this out. For now I'm using inode 9 for orphan file. I know that is reserved as EXT2_EXCLUDE_INO but at least for the sake of testing that should be fine. Honza Changes since v1: * orphan blocks have now magic numbers * split out orphan handling to a separate source file * some smaller updates according to review -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
