openshift: include network-throughput profile

[yarda]

Update network-throughput profile and drop now duplicate tuning from the openshift profile. Follows up the #809 and cleans the things a bit.

[gheffern]

I like moving these into network-throughput.

I think the addition of net.ipv4.tcp_notsent_lowat=131072 mitigates the biggest issues of increased TCP buffer sizes, those being buffer bloat latency and increased memory usage.

Because of that I almost wonder if it makes sense to reverse the relationship between throughput-performance and network-throughput. So that throughput-performance also includes network-throughput enhancements. Modern workloads tend to have a lot of network activity. Only downside I can think of is increased context switching, leading to a bit more CPU usage.

I think that would then mean these profiles would get the TCP throughput changes as well:
virtual-guest
virtual-host
accelerator-performance
oracle
mssql
sap-hana

I think its also worth updating the following comment:

# Increase kernel buffer size maximums. Currently this seems only necessary at 40Gb speeds.

To say something like:
# Increase kernel buffer size maximums to cover a wider range of bandwidth delay products on varying networks

As moderate bandwidth, higher latency links can benefit from these changes as well

[gheffern]

Thinking about it a bit more, inverting the relationship between throughput-performance and network-throughput would not be a good change for people directly depending on network-throughput already.

So if you thought it was a good idea to expand this change to those other profiles:

virtual-guest
virtual-host
accelerator-performance
oracle
mssql
sap-hana

Then we would just want to replace throughput-performance with network-throughput as you have done already for the openshift profile.

[yarda]

People who need it can also load both profiles like e.g.:

$ tuned-adm profile network-throughput oracle

Then they will have network-throughput tuning and oracle tuning together where the oracle also overrides possible conflicting tuning from the network-throughput.

[yarda]

The network tuning from #809 got upstream and into Fedora, but not yet into RHEL/CentOS. It needs approval from the performance team which may take some time. RHEL Jira issue with the customer support ticket can prioritize/speed-up the process.