Some Windows Server 2016 vs. Linux Network Benchmarks

[Guest]

A primarily Unix tool runs better on Unix, I did not see that comin'! Thanks Mike!

netperf itself is written in C. it's going to perform the same almost everywhere. Please correct me if I am wrong but aren't the interfaces pretty much the same with different implementations on different operation systems? The test here the networking performance of different operation systems. Different configuration will also affect results.

[starshipeleven]

And now we arrive at the real problems with Windows... these things are non-trivial to check for and/or control.

On Windows Server you do have more control, also on Windows Enterprise you (the sysadmin anyway) do.

[leech]

On Windows Server you do have more control, also on Windows Enterprise you (the sysadmin anyway) do.

Well, Michael is the only one who can say for sure, but it seems to me that if you're doing benchmarking you don't want to go through everything and trim things down just to get better results, you'd want to install the OS, run the benchmark at that default install. Which probably means that (hopefully!) extraneous services aren't started already. I'm just trying to figure out why Manjaro was so terrible? Bandwidth limiter on it or something?

[schmidtbag]

Maybe this should be "What was Michael thinking ...?". No mention of firewall config, no mention of CPU load before/during/after the test, ...

First of all, it's been known for a while that Michael uses whatever the distro's default settings are unless he specifies otherwise. So if you really want to know, it's not too hard to find out yourself.
Second, why would the CPU load matter when running benchmarks? When you're trying to get accurate results, you make sure the entire system is idle first; it isn't scientific or useful when you're running a performance benchmark and something is running in the background like caching or updates. So, you're nitpicking about something that no reviewers do.

[rubdos]

netperf itself is written in C. it's going to perform the same almost everywhere. Please correct me if I am wrong but aren't the interfaces pretty much the same with different implementations on different operation systems? The test here the networking performance of different operation systems. Different configuration will also affect results.

Windows indeed implements basically BSD sockets.

[starshipeleven]

Well, Michael is the only one who can say for sure, but it seems to me that if you're doing benchmarking you don't want to go through everything and trim things down just to get better results, you'd want to install the OS, run the benchmark at that default install. Which probably means that (hopefully!) extraneous services aren't started already. I'm just trying to figure out why Manjaro was so terrible? Bandwidth limiter on it or something?

The problem with using default config is that you aren't benching the actual capabilities of the hardware/software, just how good the default config is.

I understand that Micheael does not have the manpower (and skill, as optimizing stuff isn't trivial) to do more in-depth benchmarking, I just point out the limitations of the current method.

[numacross]

This benchmark is using the long term support version of Server 2016 which makes me wonder if there's much change between that and the new 1709-based version (https://docs.microsoft.com/en-us/win...rted-with-1709). It is only available in Core variant without GUI present though.

[polarathene]

Interesting how bad manjaro is doing in comparison to the others...

I'm just trying to figure out why Manjaro was so terrible? Bandwidth limiter on it or something?

Manjaro is configured for responsiveness over performance for desktop users. Changes like BFQ for I/O or increasing the kernel tick rate(1000Mhz I think), pre-emption, etc will reduce latency making things more responsive at the cost of thoroughput. CPU frequency driver and governor used might also impact results, plenty of tunables.

So I'm not all that surprised. It's a bit amusing that it's performing similar to Windows Server though which is meant to be a product that should perform better at network perf.

[nings]

The next time try to use Jumbo Frame to 4k. Like that you can reach the full 1gbps speed

[nings]

Next time, try to use Jumbo Frame to 4k like that you can probably reach the full speed of 1gbps