On Tue, 29 Mar 2022 10:46:08 +0300
Sagi Grimberg [off-list ref] wrote:

quoted

quoted

As I said, TCP can be tuned in various ways, congestion being just one
of them. I'm sure you can find a workload where rmem/wmem will make
a difference.  

agree.
but the difference for the knob of rmem/wmem is:
we could enlarge rmem/wmem for NVMe/TCP via sysctl,
and it would not bring downside to any other sockets whose
rmem/wmem are not explicitly specified.  

It can most certainly affect them, positively or negatively, depends
on the use-case.

Agree.
Your saying is rigorous.

quoted

quoted

In addition, based on my knowledge, application specific TCP level
tuning (like congestion) is not really a common thing to do. So why in
nvme-tcp?

So to me at least, it is not clear why we should add it to the driver.  

As mentioned in the commit message, though we can specify the
congestion-control of NVMe_over_TCP via sysctl or writing
'/proc/sys/net/ipv4/tcp_congestion_control', but this also
changes the congestion-control of all the future TCP sockets on
the same host that have not been explicitly assigned the
congestion-control, thus bringing potential impaction on their
performance.

For example:

A server in a data-center with the following 2 NICs:

     - NIC_fron-end, for interacting with clients through WAN
       (high latency, ms-level)

     - NIC_back-end, for interacting with NVMe/TCP target through LAN
       (low latency, ECN-enabled, ideal for dctcp)

This server interacts with clients (handling requests) via the fron-end
network and accesses the NVMe/TCP storage via the back-end network.
This is a normal use case, right?

For the client devices, we can’t determine their congestion-control.
But normally it’s cubic by default (per the CONFIG_DEFAULT_TCP_CONG).
So if we change the default congestion control on the server to dctcp
on behalf of the NVMe/TCP traffic of the LAN side, it could at the
same time change the congestion-control of the front-end sockets
to dctcp while the congestion-control of the client-side is cubic.
So this is an unexpected scenario.

In addition, distributed storage products like the following also have
the above problem:

     - The product consists of a cluster of servers.

     - Each server serves clients via its front-end NIC
      (WAN, high latency).

     - All servers interact with each other via NVMe/TCP via back-end NIC
      (LAN, low latency, ECN-enabled, ideal for dctcp).  

Separate networks are still not application (nvme-tcp) specific and as
mentioned, we have a way to control that. IMO, this still does not
qualify as solid justification to add this to nvme-tcp.

What do others think?

Well, per the fact that the approach (‘ip route …’) proposed
by Jakub could largely fit the per link requirement on
congestion-control, so the usefulness of this patchset is really
not so significant.

So here I terminate all the threads of this patchset.

At last, many thanks to all of you for reviewing this patchset.