Re: Adding xxhash to gluster code base

[Aravinda <avishwan@xxxxxxxxxx>]

regards
Aravinda VK

On 06/27/2017 01:38 PM, Niels de Vos wrote:

On Tue, Jun 27, 2017 at 12:25:11PM +0530, Kotresh Hiremath Ravishankar wrote:

Hi,

We were looking for faster non-cryptographic hash to be used for the
gfid2path infra [1]
The initial testing was done with md5 128bit checksum which was a slow,
cryptographic hash
and using it makes software not complaint to FIPS [2]

On searching online a bit we found out xxhash [3] seems to be faster from
the results of
benchmark tests shared and lot of projects use it. So we have decided to us
xxHash
and added following files to gluster code base with the patch [4]

    BSD 2-Clause License:
       contrib/xxhash/xxhash.c
       contrib/xxhash/xxhash.h

    GPL v2 License:
       tests/utils/xxhsum.c

NOTE: We have ignored the code guideline check for these files as
maintaining it
further becomes difficult.

Please comment on the same if there are any issues around it.

How performance critical is the hashing for gfid2path? 

For this we need non-crypto hashing. This hash calculation will be in the I/O path(Affected FOPs are create, mknod, link, symlink, rename, unlink)

What is the plan to keep these files maintained? At minimal we need to
add these files to MAINTAINERS and the maintainers need to cherry-pick
updates and bugfixes from the original project. The few patches a year
makes this a recurring task that should not be forgoten. It would be
much better to use this as an external library that is provided by the
distributions. We already rely on OpenSSL, does this library not provide
an alternative 'FIPS approved' hashing that performs reasonably well?

I think, this library is not available in distribution packaging yet. Please suggest if you know any non-crypto hashing lib which is already available in distros.

Some distributions are very strict on bundling external projects, and we
need to inform the packagers about the additions so that they can handle
it correctly. Adding an external project to contrib/ should be mentioned
in the release notes at the very least.

Note that none of the symbols of any public functions in Gluster may
collide with functions in standard distribution libraries. This causes
for regular problems with gfapi applications. All exposed symbols that
get imported in contrib/ should have a gf_ prefix.

Thanks,
Niels

[1] Issue: https://github.com/gluster/glusterfs/issues/139
[2] https://en.wikipedia.org/wiki/Federal_Information_Processing_Standards
[3] http://cyan4973.github.io/xxHash/
[4] https://review.gluster.org/#/c/17488/10



-- 
Thanks and Regards,
Kotresh H R and Aravinda VK

_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://lists.gluster.org/mailman/listinfo/gluster-devel

_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://lists.gluster.org/mailman/listinfo/gluster-devel