Hi, I've posted an initial implementation of a new pool feature called large_dnode for review. The Linux and Illumos ports are posted here:
https://github.com/zfsonlinux/zfs/pull/3542 https://reviews.csiden.org/r/231/ This work is based on a prototype originally developed by Prakash Surya. I have extended that prototype to support larger bonus buffers to take advantage of the increased dnode size. The goal is to avoid the use of spill blocks in order to increase performance of metadata operations, reduce memory overhead, and to reduce pool fragmentation. More detailed notes can be found in the commit message on github or in the review request description. This is mainly useful today on ZFS on Linux systems that make heavy use of extended attributes, such as Lustre servers and systems using SELinux. On other platforms, the only use cases I'm aware of that may benefit from this today include large ACLs or long symbolic link target names. This feature may be desirable on non-Linux platforms in the future if applications or features are developed that could make use of a larger bonus buffer. I am very interested in getting feedback, testing, and buy-in from the OpenZFS community before we consider merging this work in the Linux port. I wasn't able to build and test on Illumos due to my lack of familiarity with that environment. If any Illumos folks have time to try it out I would be very grateful. Thanks in advance for your feedback. Ned _______________________________________________ developer mailing list [email protected] http://lists.open-zfs.org/mailman/listinfo/developer
