Re: [bmwg] Review of draft-dcn-bmwg-containerized-infra-10

[Minh Ngoc Tran <[e-mail]>]

Hi Vratko and Gábor,

This is the second part of the reply


*Regarding Vratko's review about the Resource Configuration part:*

Ok, that is enough about scope.

I have few ideas for additions to Resources configuration (section 4.2).



1. Some resources also apply to VM-based VNFs. It would be nice to mention
which and how.



Will be updated as suggestion about NUMA and CPU Isolation. hugepages and
CPU/memory allocation here are specific to pod.



2. Mention the current list of resources is not exhaustive.

If a deployment is affected by some additional resources,

the report should include them (an their configuration).



As this document is not aim for benchmarking standard. Benchmark reporting
is out of scope of this document.



3. SUT is composed of NICs, (possibly cables between hosts),

kernel and/or vSwitch, C-NFVs, and various connections between components

(shared memory, PCI, both perhaps cross-NUMA,

or on the other hand both perhaps accessing cache instead of main memory).

While the draft usually considers the NFVs to be the DUTs,

most of the considerations also applies for benchmarking other SUT
components.



Will add this suggestion at the beginning of the Resource Configuration
section



4. Mention that CPU pinning and other isolation practices not only

reduce the performance variation and degradation caused by the “noisy
neighbor” problem

(VNFs are neighboring both other NFVs and vSwitch),

it also usually “bottlenecks” the overall SUT behavior.

That means sometimes NVFs are the bottleneck (vSwitch performance does not
matter),

sometimes the vSwitch is the bottleneck (e.g. when NFV is just testpmd as
in appendix C).



Will add this suggestion in CPU Isolation sub-section



5. Draft may even recommend varying non-DUT resources to prove

the intended DUT is really the bottleneck

(or at least has the highest impact on SUT performance).

Sometimes varying some resource is not possible,

and sometimes the isolation is not perfect enough.

(MLRsearch will help with overall performance variation

but will not verify isolation nor bottleneck.)



Could you explain about the non-DUT resources?

Because the SUT already consists of C-VNFs, CNI, vSwitch, NIC.

--------------


Besides these changes corresponding to the review,

we will update the benchmarking appendixes with latest result and re-orgranize
them to match with the considerations in the document

The revised benchmarking appendixes in our next draft version will have the
following section

- Benchmarking Experience (Different Networking Models -
userspace/ebpf/smart-nic)

- Benchmarking Experience (Resources Configuration in Single Pod Scenario)

- Benchmarking Experience (Networking Model Combination and Resources
Configuration in Multi-Pod Scenario)



Once again, Thank you very much for reading and commenting on our draft !
Best regards,

Ngoc.

__________________________________

*Minh-Ngoc Tran*

*Distributed Cloud and Network lab*

*Soongsil University *

*511 Sangdo-dong Dongjak-gu*

*Seoul, 156-743 Korea*

*TEL : [phone] *