[Gluster-devel] [Gluster-Maintainers] Release 5: Performance comparisons

[Shyam Ranganathan]

On 09/26/2018 10:21 AM, Shyam Ranganathan wrote:
> 4. Performance testing/benchmarking
>   - I would be using smallfile and FIO to baseline 3.12 and 4.1 and test
> RC0 for any major regressions
>   - If we already know of any please shout out so that we are aware of
> the problems and upcoming fixes to the same


Managed to complete this yesterday, and attached are the results. The
comparison is between 4.1.5 and 5.0 to understand if there are any mojor
regressions, the tests themselves need to be tuned better in the future,
but help to provide an initial look into the comparing releases.

Observations (just some top ones, not details, considering at least a 5%
delta from prior release numbers):
- ls -l tests with smallfile on dist-arbiter volumes seems to have regressed
- create tests with smallfile on dist-arbiter volumes seems to have improved
- FIO sequential write performance remains mostly the same across volume
types
- FIO sequential read performance seems to degrade on certain volume types
- FIO random write performance seems to degrade on certain volume types
- FIO random read performance seems to have improved across volume types

Goof-ups:
- The volume creation ansible play just laid out bricks in host order,
hence for tests like dist-dispers-4x(4+2) all 6 bricks of the same
subvolume ended up on the same host. Interestingly this happened across
both versions compared, and hence the pattern was the same, allowing
some base comparisons.

Next steps:
- I will be running the tests that gave inconsistent results with RC1
when we build the same
- It would be useful for component owners to take a look and call out
possible causes for some of the degrades, if already known

Shyam
-------------- next part --------------
A non-text attachment was scrubbed...
Name: gbench-Summary-4.1.5-to-5.0.ods
Type: application/vnd.oasis.opendocument.spreadsheet
Size: 23340 bytes
Desc: not available
URL: <http://lists.gluster.org/pipermail/gluster-devel/attachments/20181002/ab04578d/attachment-0001.ods>