Managed RIO Server

[benaadams]

Updated commit for #10

[DamianEdwards]

Latest results below. Seems CPU use went up a bit and we lost peak RPS (down from >7m) as a result.

Connections	Pipeline Depth	Bandwidth	RPS	CPU
1	1	25 Mbps	~14k	13%
8	16	1.1 Gbps	~855k	37%
8	512	7.3 Gbps	~5.4m	40%
32	16	2.9 Gbps	~2.2m	80%
128	16	6.0 Gbps	~4.5m	100%
256	16	6.8 Gbps	~5.1m	100%
512	16	6.9 Gbps	~5.2m	100%
512	150	8.8 Gbps	~6.7m	84%
512	512	9.1 Gbps	~6.9m	67%

[dnfclas]

Hi @benaadams, I'm your friendly neighborhood .NET Foundation Pull Request Bot (You can call me DNFBOT). Thanks for your contribution!
You've already signed the contribution license agreement. Thanks!

The agreement was validated by .NET Foundation and real humans are currently evaluating your PR.

TTYL, DNFBOT;

[Eilon]

@DamianEdwards CLA bot is happy.

[benaadams]

Hopefully this will max network before CPU

[DamianEdwards]

Looking forward to trying this one 😄

[benaadams]

Tried it on an Azure G5; need to put some SIMD fixes in as on that hardware Vector.Count operates on 32 bytes at the same time rather than 16 bytes (AVX 256 rather than SSE 128)

[benaadams]

wrk -c 512 --latency -t 32 http://url/ -s ./scripts/pipeline.lua -- 148

Still a bit funny on NUMA - good but not great, probably need to do thread affinities there; but on Azure G4: 9.0 Gbps at 59% CPU:

At 6.8M requests per second (6,807,542). Latency of 92.ms (avg), 113.81 ms (std-dev), 1.27s (max), 50% < 43ms, 90% < 288ms.

[Bobris]

You do all crazy stuff with AVX and then use this in performance critical code???

                var date = DateTime.UtcNow.ToString("r");
                Encoding.UTF8.GetBytes(date, 0, dateBytes.Length, dateBytes, 0);

[benaadams]

You do all crazy stuff with AVX and then use this in performance critical code???

😉

Lazy fix, function could be faster, but still should cut down the string allocs by 6.9M per second at 9.1 Gbps, to 512 per second; so not a bad change...

Takes my test on 68% CPU to 61%, so seems effective.

[Bobris]

I think you would found DateTime.UtcNow slow too (it is basically Kernel call). This is how simply I do it in Nowin:
https://github.com/Bobris/Nowin/blob/master/Nowin/DateHeaderValueProvider.cs

[benaadams]

This is how simply I do it in Nowin

That's much better than the convoluted improvement I was planning; gone for something similar.

[RichiCoder1]

No dev comment, just wanted to say awesome job @benaadams 👍

[CIPop]

This is really cool.
@benaadams it would be very interesting to see how this compares with a native-only implementation of RIO: http://ctstraffic.codeplex.com/

[benaadams]

@CIPop A reference implementation!

Super useful to know; I couldn't really find anything other than the api docs building this so its mostly trial and error iteration; will definitely have a look at it; thank you 👍

[CIPop]

Glad to help :)
I just learned about this yesterday. I'll update the Registered I/O Extensions tracking issue in CoreFX as well.

[davidfowl]

Closing this out