Hi. In general, I agree that the code could use a more rigid and
easier-to-follow structure. Please see this email I sent last week to
the ovs-dev list:
https://mail.openvswitch.org/pipermail/ovs-dev/2021-January/379774.html
I'll address some of your points below.
On 2/2/21 6:35 AM, gongming chen wrote:
To be honest, I think the direction of the ovn community is not towards
a stable and clear structure. Especially in the patch integration in the
past year, the new patch does not take too much into consideration for
compatibility.Many of them are completely logical changes, so the cost
of learning and following the ovn community is greatly increased. And
many logical changes are completely unnecessary.
I disagree here. Every patch series submitted for OVN either fixes a
bug, adds a new feature, or fixes a performance bottleneck. Some
individual patches within a series might reorganize code in order to
make later patches in the same series easier to follow, though.
For example, ofctrl_dup_flow--->ovn_flow_dup, when looking at the new
github code, I found that ofctrl_dup_flow was not found, and the new
function did not change anything except the name. I think this change is
completely unnecessary, except Will increase the time cost of reading
new code.
It also changed the function to be callable only from within ofctrl.c.
The name change helps to hint that it is not part of the public ofctrl_*
API. It's also moot since ovn_flow_dup() was removed in a later patch in
the same series [1].
Complexity such as (no other maliciousness), Incremental processing for
flow installation by tracking:
Compared with the small benefits it brings (ofctl put itself is not a
big bottleneck), I think the complexity it brings, and the changes
brought by the code structure (the change of the existing version is a
big cost) Is bigger. This is not conducive to the long-term stable
development of ovn, because the code structure of ovn is becoming
increasingly unclear.
I agree with the final part of this: the structure of ovn is becoming
increasingly unclear. It is harder to make new changes to ovn without
causing regressions, because a change in one area of the code has the
potential to affect a seemingly unrelated area.
However, I will contend that incremental processing for flow processing
was a large bottleneck in ovn-controller. According to the cover letter
for that patch series
(https://mail.openvswitch.org/pipermail/ovs-dev/2020-August/374099.html), adding
incremental processing dropped CPU usage by 39% for the port binding
operation on a 1200 HV, 12K lport network. This is in keeping with
previous profiling results that had shown that reconciling desired and
installed flows was the second highest consumer of CPU in ovn-controller
after lflow expression parsing [2].
As a general community participant, with long-term ovn production
environment operating experience, I think the big bottleneck of ovn now
is stability and structural clarity (for custom development). I hope ovn
has clear code (like the code structure of the kernel), which can reduce
the time cost of using it. I hope ovn is small, clear, stable, and
reliable. Some personal suggestions, and finally hope that the ovn
community will get better and better.
Best wishes.
Thanks, I'm in agreement that the code needs clearer structure. The
hardest of those descriptors to deliver on is "small". Users are
demanding larger and more complex scenarios than when ovn was initially
developed. With those scenarios, we are finding that delivering the
requested performance has come at the cost of code complexity. However,
the suggested refactor for OVN is made with the idea of better defining
the structure of the code so that it is easier to document, follow, and
can be more easily tested.
Mark Michelson
_______________________________________________
discuss mailing list
disc...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-discuss
[1] You may be asking "Why would this function be added, only to be
removed later in the same patch series?" That's due to the review
process. The first version of the series changed the function name to
ovn_flow_dup() and left it that way for the entire series. In code
review, it was suggested to split desired and installed flows into
separate types. The result is that a later version of the patch series
created ovn_flow_dup() in patch 1 and then changed it to
installed_flow_dup() in patch 6.
[2] lflow_expression_parsing is avoided when possible due to incremental
processing of logical flows, as well as caching expression results when
possible. Both of these changes fall into the category of "increases
performance but makes the code more complex".
_______________________________________________
discuss mailing list
disc...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-discuss