High differences between run for GC Diagnoser

[xoofx]

Trying to play with the GCDiagnoser and I'm getting weird results on consecutive run, like the Gen0/Gen1/Gen2/Bytes Allocated/Op are varying a lot (30-40%), making it quite useless to compare... while I would not expect figures to change that much (afaik, I'm not aware about this indeterministic GC behaviour)

Have you been noticing this issue or Am I doing something wrong there?

[adamsitnik]

I have been running our integration tests plenty of times, they contain simple test for GC Diagnoser and the results were always the same.

.NET GC is selftunning (changing Gen sizes etc) but we always start new process so the results should be almost the same as long as OS has similar resources available.

Could you give some sample benchmark with config that you use?

[mattwarren]

Can you post the logs from 2 runs with different # of GC collections? It'll be easier to see what's going on

On 1 Apr 2016, at 02:48, "Alexandre Mutel" notifications@github.com wrote:

Trying to play with the GCDiagnoser and I'm getting weird results on consecutive run, like the Gen0/Gen1/Gen2/Bytes Allocated/Op are varying a lot (30-40%), making it quite useless to compare... while I would not expect figures to change that much (afaik, I'm not aware about this indeterministic GC behaviour)

Have you been noticing this issue or Am I doing something wrong there?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub

[xoofx]

The output I'm getting while running multiple times is like this:

               Method |    Median |    StdDev |  Gen 0 | Gen 1 | Gen 2 | Bytes Allocated/Op |
--------------------- |---------- |---------- |------- |------ |------ |------------------- |
    TestCommonMarkNet | 4.8020 ms | 0.0326 ms | 208.54 | 14.34 | 89.28 |       1,117,194.48 |
 TestCommonMarkNetNew | 5.7072 ms | 0.0541 ms | 169.87 | 84.31 | 72.44 |       1,124,892.65 |
          TestMarkdig | 6.4552 ms | 0.0454 ms | 115.00 | 75.00 | 55.00 |         851,150.77 |
     TestMarkdownDeep | 8.4201 ms | 0.5206 ms | 176.43 | 76.75 | 56.46 |       1,039,481.68 |

               Method |    Median |    StdDev |  Gen 0 | Gen 1 | Gen 2 | Bytes Allocated/Op |
--------------------- |---------- |---------- |------- |------ |------ |------------------- |
    TestCommonMarkNet | 4.7802 ms | 0.0365 ms | 117.28 |  8.10 | 50.15 |         738,831.91 |
 TestCommonMarkNetNew | 5.7256 ms | 0.0317 ms | 176.00 | 87.42 | 75.26 |       1,371,935.85 |
          TestMarkdig | 6.2702 ms | 0.0673 ms |  91.59 | 59.74 | 43.81 |         797,976.80 |
     TestMarkdownDeep | 7.9212 ms | 0.0405 ms | 164.00 | 71.00 | 52.00 |       1,134,079.22 |


               Method |    Median |    StdDev |  Gen 0 | Gen 1 | Gen 2 | Bytes Allocated/Op |
--------------------- |---------- |---------- |------- |------ |------ |------------------- |
    TestCommonMarkNet | 4.8237 ms | 0.0504 ms | 152.82 | 10.51 | 65.42 |       1,117,194.48 |
 TestCommonMarkNetNew | 5.6900 ms | 0.0490 ms | 138.97 | 68.97 | 59.27 |       1,255,915.01 |
          TestMarkdig | 6.3441 ms | 0.0963 ms |  70.75 | 46.02 | 33.66 |         713,802.94 |
     TestMarkdownDeep | 7.8991 ms | 0.0547 ms | 164.00 | 71.00 | 52.00 |       1,314,925.80 |    


               Method |    Median |    StdDev |  Gen 0 |  Gen 1 |  Gen 2 | Bytes Allocated/Op |
--------------------- |---------- |---------- |------- |------- |------- |------------------- |
    TestCommonMarkNet | 4.7855 ms | 0.0920 ms | 194.88 |  13.53 |  83.23 |         963,242.75 |
 TestCommonMarkNetNew | 5.7061 ms | 0.0456 ms | 249.01 | 123.83 | 106.91 |       1,522,547.54 |
          TestMarkdig | 6.3259 ms | 0.1434 ms | 115.00 |  75.00 |  55.00 |         785,682.84 |
     TestMarkdownDeep | 7.9409 ms | 0.0626 ms | 173.00 |  75.00 |  55.00 |         938,949.44 |

               Method |    Median |    StdDev |  Gen 0 |  Gen 1 | Gen 2 | Bytes Allocated/Op |
--------------------- |---------- |---------- |------- |------- |------ |------------------- |
    TestCommonMarkNet | 4.7873 ms | 0.0881 ms | 188.43 |  13.16 | 80.36 |         992,620.58 |
 TestCommonMarkNetNew | 5.7098 ms | 0.0235 ms | 212.03 | 105.35 | 90.77 |       1,381,368.58 |
          TestMarkdig | 6.4275 ms | 0.0694 ms | 113.54 |  74.13 | 54.42 |         827,121.71 |
     TestMarkdownDeep | 7.9162 ms | 0.0472 ms | 173.00 |  75.00 | 55.00 |       1,000,627.30 |

As you can see, sometimes figures are jolting for the Gem0/Gen1/Gen2 (worse for Bytes Allocated/Op), changing even the ranking between runs.

When I'm running for example a simple benchmark (without using BenchmarkDotNet) and outputing GC.CollectionCount(0)/GC.CollectionCount(1)/GC.CollectionCount(2)) and testing one solution (like TestcommontMarkNet), the GC counts are fixed and not changing between runs.

[xoofx]

Looking at the code, one thing I'm not sure is that I would expect the GCDiagnoser to only process GC events for the IterationMode.MainTarget period: Typically there are other periods that seem to be non deterministic (like the IdleTarget, not sure what it does, but it looks like it is adapting to the process) and we don't want them as part of the measurement. Also, not sure that it is good to have variable #operation count for different method when measuring GC events...

[xoofx]

Ok, so after trying to tweak the usage of BenchmarkDotNet and using Mode.SingleRun which disable the Pilot phase, numbers are completely stable (and much more equivalent between the methods).

Though, not sure I did use correctly the API, but I had to patch the MethodInvoker.Invoke method to hardcode invokeCount = 64, because if I tried to setup the attribute on each method [Benchmark(OperationsPerInvoke = 64)], the methods are actually not executed and gives completely invalid results... so difficult to understand what's going on with this...

I used the following config:

var config = ManualConfig.Create(DefaultConfig.Instance);
var gcDiagnoser = new GCDiagnoser();
config.Add(new Job { Mode = Mode.SingleRun, LaunchCount = 2, WarmupCount = 2, TargetCount = 10 });
config.Add(gcDiagnoser);

[mattwarren]

As you can see, sometimes figures are jolting for the Gem0/Gen1/Gen2 (worse for Bytes Allocated/Op), changing even the ranking between runs

The Gen 0/1/2 values are scaled/normalised based on the min value across the benchmarks (see https://github.com/PerfDotNet/BenchmarkDotNet/blob/master/BenchmarkDotNet.Diagnostics/GCDiagnoser.cs#L174), so it's not surprising they vary as the # of runs/Ops changes.

However the # Bytes is per Op, so should be more stable

[mattwarren]

Yeah 'OperationsPerInvoke' isn't going to help, that's for something else.
However you should be able to use 'IterationCount' on the Job, see
https://github.com/PerfDotNet/BenchmarkDotNet/blob/master/BenchmarkDotNet/Jobs/Job.cs#L28

Looking at the code, one thing I'm not sure is that I would expect the GCDiagnoser to only process GC events for the IterationMode.MainTarget period: Typically there are other periods that seem to be non deterministic (like the IdleTarget, not sure what it does, but it looks like it is adapting to the process) and we don't want them as part of the measurement. Also, not sure that it is good to have variable #operation count for different method when measuring GC event

Good point, I need to have a think about how to best handle this.

[xoofx]

Setting IterationTime doesn't work for me:

var config = ManualConfig.Create(DefaultConfig.Instance);
var gcDiagnoser = new GCDiagnoser();
config.Add(new Job { Mode = Mode.SingleRun, LaunchCount = 2, WarmupCount = 2, IterationTime = 64, TargetCount = 10 });
config.Add(gcDiagnoser);

It results in only 1 op:

MainTarget 1: 1 op, 4785568.28 ns, 4.7856 ms/op
MainTarget 2: 1 op, 4660543.99 ns, 4.6605 ms/op

[xoofx]

Looking at MethodInvoker.cs:92 , IterationTime is only used for the Throughput mode and not the SingleRun

[xoofx]

Also on a side note, I had to compile BenchmarkDotNet.Diagnostics on my side as it is not published to nuget. Is it an oversight?

[xoofx]

Hm, ok, maybe Diagnostics cannot be released because xproj+project.json doesn't support yet embedding native files... arg... It seems we will have to wait for dotnet/cli :(

[mattwarren]

Hm, ok, maybe Diagnostics cannot be released because xproj+project.json doesn't support yet embedding native files... arg... It seems we will have to wait for dotnet/cli :(

Hmm, didn't realise that @adamsitnik or @AndreyAkinshin does this block us from releasing a BenchmarkDotNet.Diagnostics nuget package?

[mattwarren]

@xoofx are you happy with the workaround you have, i.e. using a custom Config?

Or would you want BenchmarkDotNet to be more stable/consistent across runs?

[xoofx]

Hmm, didn't realise that @adamsitnik or @AndreyAkinshin does this block us from releasing a BenchmarkDotNet.Diagnostics nuget package?

Actually, I was able to ship native DLLs packInclude in project.json, along with a msbuild .targets file for net20-net4x+ platforms, so that is finally possible at least.

Or would you want BenchmarkDotNet to be more stable/consistent across runs?

Hm, actually, no as I said above, IterationTime is not working, so I still had to patch the BenchmarkDotNet.
At least, when using GCDiagnoser, it is only useful to run it in a stable mode, otherwise figures are completely useless.
I'm wondering if it is really good to have this adaptive Iteration count done for each method... I would expect that the first run returns its adaptive results and other methods measured would use the exact same iteration. But I understand that in scenarios where figures are changing a lot between the methods, that would be not ideal (if a very fast method is used as a baseline for the count, and then a much slower method is used, it would take more time to run it...)... so yeah, not sure how to deal with this