Reduced cores from ~40k to ~20k by moving to .NET Core | Hacker News Reader

Reduced cores from ~40k to ~20k by moving to .NET Core

devblogs.microsoft.com

30 pointsbenaadams13 comments

.Net has shown some impressive performance improvements with .Net core. Double the performance of already performent runtime is commandable.

.net is now much faster tha Java on all but one benchmark https://benchmarksgame-team.pages.debian.net/benchmarksgame/...

.NET has made tremendous strides in performance, kudos to the team.

That being said, some of the results of those benchmarks do not hold on modern CPUs. Try to run the binarytrees benchmark for example (especially in a loop) and you'll see that the JVM is quicker than .NET, especially after JIT kicks in.

Please show what you measured and how you measured.

Java (openjdk 15.0.1 2020-10-20) $ java binarytrees 21

    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 2.640825365s
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 1.856262182s
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 1.650234063s

C# (.NET Core 5.0.100)

    $ dotnet run -c Release 21
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 00:00:09.2958748
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 00:00:09.8271153
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 00:00:09.8948142

The benchmarks game shows 2.5 elapsed secs for Java and 2.4 for C#

You seem to be showing 2.6s for Java but 9.9s for C# ?

Did you use ServerGarbageCollection for C# ?

> Did you use ServerGarbageCollection for C# ?

I have not. Thanks for pointing it out.

Now that I set it, it's quite a bit faster, though not as fast as Java:

    $ java binarytrees 21
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 2.841672099s
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 1.653618362s
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 1.471093907s
    
    $ dotnet ./bin/Release/net5.0/osx-x64/binarytree.dll 21
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 00:00:01.8994677
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 00:00:01.8563705
    
    stretch tree of depth 22  check: 8388607
    2097152  trees of depth 4  check: 65011712
    524288  trees of depth 6  check: 66584576
    131072  trees of depth 8  check: 66977792
    32768  trees of depth 10  check: 67076096
    8192  trees of depth 12  check: 67100672
    2048  trees of depth 14  check: 67106816
    512  trees of depth 16  check: 67108352
    128  trees of depth 18  check: 67108736
    32  trees of depth 20  check: 67108832
    long lived tree of depth 21  check: 4194303
    
    Time taken: 00:00:01.7578978

Thanks for re-measuring.

You started off saying — "some of the results of those benchmarks do not hold on modern CPUs" — but the first Java time you measure is actually slower than the time shown on the benchmarks game?

Please consider using JMH to make your repeated Java measurements:

https://github.com/openjdk/jmh/

Yes, I think there was something else running in the background at the same time.

JMH is definitely the way to go.

Note how the JVM JIT kicks in from the 2nd iteration onward. This was tested on a 2019 Macbook pro 2.6GHz i7. I took the fastest Java and C# submissions.

What code changes have you made to make repeat timings? Or did you use JMH ?

Earlier bing also moved from .net Framework to .net core and saw 34% improvement. https://devblogs.microsoft.com/dotnet/bing-com-runs-on-net-c...

When they say 40k cores do they really mean 40,000 cores? Or is that some type of kube measurement system I'm not familiar with

Reply on news.ycombinator.com