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

[Gabor LENCSE <[e-mail]>]

Dear Authors and BMWG,

I have read the main text of draft-dcn-bmwg-containerized-infra-10 and 
quickly gone through its appendix, too.

My impression is that this is a useful document and gives important 
insight of what aspects to consider when benchmarking various parts of 
containerized infrastructure.

Special thanks for Section 3. It gives an essential introduction for 
those who do bechmarking but not really engaged in various 
containerization solutions. I think you could add some more details and 
you should definitely add references. (It uses a lot of technical terms, 
vendor and product names, etc. without references. E.g., SRIOV, eBPF, 
MACVLAN/IPVLAN,  Docker, CRI-O, containerd, Calicom, Flannel, just to 
mention a few.) I saw that references were later provided for some of 
them, but they should be provided at the first occurrence.

I cannot judge the correctness and/or completeness of the information in 
Section 4 as I am not an expert of the field. At least it seems to be 
reasonable to me. :-)

A few critical comments:

Section 4.2.2 Hugepages

In the containerized infrastructure,
    the container is isolated at the application level, and
    administrators can set huge pages more granular level (e.g.,
    Kubernetes allows to use of*512M bytes huge pages*  for the container
    as default values).

I do not know Kubernets, but I use hugepages with DPDK. Intel CPUs 
usually support only two sizes: 2MB and 1GB. Current practice is to use 
1GB hugepage size, 2MB is just for very old CPUs that do not support 1GB.


In Figure 12: Test Results:

           +--------------------+---------------------+-------------+
           |        Model       |  NUMA Mode (pinning)| Result(Gbps)|
           +--------------------+---------------------+-------------+
           |                    |          N/A        |     3.1     |
           |  Maximum Line Rate |---------------------+-------------+
           |                    |      same NUMA      |     9.8     |
           +--------------------+---------------------+-------------+

However, I did not find the applied frame size, thus regarding the 
various line rate in Gbps, it is not clear whether they were achieved 
with what frame size (e.g., 64 byte or 1518 byte).


I have not found mentioning any IP version in the draft. I suspect that 
IPv4 was used for all tests. However, I think that since 2017, the 
publication date of RFC 8200, IPv6 should not be just neglected.


All in all, I think this draft is useful and I will support its 
adoption. I am aware that saying so implies the implicit promise of 
another, more through review in the future. :-)

Best regards,

Gábor