On 12-Jun-18 2:07 PM, Thomas Monjalon wrote:
+Cc Jananee
07/06/2018 23:01, Anatoly Burakov:
Previously, unit tests were running in groups. There were
technical reasons why that was the case (mostly having to do
with limiting memory), but it was hard to maintain and update
the autotest script.
In 18.05, limiting of memory at DPDK startup was no longer
necessary, as DPDK allocates memory at runtime as needed. This
has the implication that the old test grouping can now be
retired and replaced with a more sensible way of running unit
tests (using multiprocessing pool of workers and a queue of
tests). This patchset accomplishes exactly that.
This patchset conflicts with some of the earlier work on
autotests [1] [2] [3], but i think it presents a cleaner
solution for some of the problems highlighted by those patch
series. I can integrate those patches into this series if
need be.
[1] http://dpdk.org/dev/patchwork/patch/40370/
[2] http://dpdk.org/dev/patchwork/patch/40371/
[3] http://dpdk.org/dev/patchwork/patch/40372/
It may be interesting to work on lists of tests as done
in the following patch:
http://dpdk.org/dev/patchwork/patch/40373/
The idea is to split tests in several categories:
- basic and short test
- longer and lower priority
- performance test
As a long term solution, we can think about making category an attribute
inside the test itself?
These test categories do not conflict with my patchset as they
ultimately rely on white/blacklisting, which will continue to work as
before.
In my view, it really boils down to two things - either tests can be run
in parallel with others (i.e. their result won't be affected by another
independent DPDK test app instance), or not. On top of that, we can use
blacklisting or whitelisting to define which tests will actually be run
(i.e. define any "categories" we want), but their (non-)parallelism must
always be respected to get good test results.
--
Thanks,
Anatoly
