Re: [rng] Usefulness of benchmarks
Hi. On Sun, 4 Sep 2016 15:20:39 +0300, Artem Barger wrote: On Sun, Sep 4, 2016 at 3:01 PM, Gilleswrote: "RandomStressTester" is in Commons _Rng_ "src/userguide" example code, while I've just committed "RandomNumberGeneratorBenchmark" in Commons _Math_ userguide examples (branch "develop"). There is a README in "src/userguide"; basically, you'd run with a command similar to: $ mvn -q exec:java \ -Dexec.mainClass=org.apache.commons.math4.userguide.rng.RandomNumberGeneratorBenchmark \ -Dexec.args="MT ISAAC JDK" where the "exec.args" contains a list of "RandomSource" identifiers. I've got following results and they look similar to ones you have in the table :/ [...] It seems that the behaviour changes with the number of generators being tested. I don't know what to make out of that... Personally, I think that these codes are so small/fast (wrt to common use of Java) that such micro-benchmarking measurements border to an exercise in futility. Concerning the subject of this thread, I guess that we are back to square one, i.e. no more knowledge about it was gathered here than we had in December 2015, even though back then the benchmark score was the sole argument put forward for constraining the design. :-( Regards, Gilles - To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
Re: [rng] Usefulness of benchmarks
On Sun, Sep 4, 2016 at 3:01 PM, Gilleswrote: > "RandomStressTester" is in Commons _Rng_ "src/userguide" example > code, while I've just committed "RandomNumberGeneratorBenchmark" > in Commons _Math_ userguide examples (branch "develop"). > > There is a README in "src/userguide"; basically, you'd run with a > command similar to: > > $ mvn -q exec:java \ > > -Dexec.mainClass=org.apache.commons.math4.userguide.rng.RandomNumberGeneratorBenchmark > \ >-Dexec.args="MT ISAAC JDK" > > where the "exec.args" contains a list of "RandomSource" identifiers. > > I've got following results and they look similar to ones you have in the table :/ Java 1.8.0_101 Runtime 1.8.0_101-b13 JVM Java HotSpot(TM) 64-Bit Server VM 25.101-b13 nextInt() (calls per timed block: 100, timed blocks: 500, time unit: ms) name time/call std dev total time ratio cv difference j.u.Random 1.390e-05 4.2e-06 6.9486e+03 1.000 0.30 0.e+00 o.a.c.r.i.s.MersenneTwister 1.129e-05 3.9e-06 5.6437e+03 0.812 0.34 -1.3049e+03 o.a.c.r.i.s.ISAACRandom 1.611e-05 3.7e-06 8.0537e+03 1.159 0.23 1.1052e+03 o.a.c.r.i.s.JDKRandom 1.634e-05 3.4e-06 8.1704e+03 1.176 0.21 1.2218e+03 nextDouble() (calls per timed block: 100, timed blocks: 500, time unit: ms) name time/call std dev total time ratio cv difference j.u.Random 2.608e-05 4.3e-06 1.3042e+04 1.000 0.17 0.e+00 o.a.c.r.i.s.MersenneTwister 2.099e-05 3.9e-06 1.0497e+04 0.805 0.19 -2.5450e+03 o.a.c.r.i.s.ISAACRandom 2.381e-05 3.8e-06 1.1905e+04 0.913 0.16 -1.1367e+03 o.a.c.r.i.s.JDKRandom 2.922e-05 4.5e-06 1.4612e+04 1.120 0.15 1.5705e+03 nextLong() (calls per timed block: 100, timed blocks: 500, time unit: ms) name time/call std dev total time ratio cv difference j.u.Random 2.436e-05 3.9e-06 1.2180e+04 1.000 0.16 0.e+00 o.a.c.r.i.s.MersenneTwister 1.890e-05 3.4e-06 9.4501e+03 0.776 0.18 -2.7297e+03 o.a.c.r.i.s.ISAACRandom 2.169e-05 4.1e-06 1.0845e+04 0.890 0.19 -1.3352e+03 o.a.c.r.i.s.JDKRandom 2.743e-05 4.7e-06 1.3713e+04 1.126 0.17 1.5330e+03 Best regards, Artem Barger.
Re: [rng] Usefulness of benchmarks
On Sun, 4 Sep 2016 14:22:04 +0300, Artem Barger wrote: On Sun, Sep 4, 2016 at 2:06 PM, Gilleswrote: A few quick runs seems to produce numbers that are now much closer to what JMH produces.[1] What do you think? How I can produce results with PerfTestUtils? Is there any script to run RandomStressTester? "RandomStressTester" is in Commons _Rng_ "src/userguide" example code, while I've just committed "RandomNumberGeneratorBenchmark" in Commons _Math_ userguide examples (branch "develop"). There is a README in "src/userguide"; basically, you'd run with a command similar to: $ mvn -q exec:java \ -Dexec.mainClass=org.apache.commons.math4.userguide.rng.RandomNumberGeneratorBenchmark \ -Dexec.args="MT ISAAC JDK" where the "exec.args" contains a list of "RandomSource" identifiers. It possible that previous results were produced on loaded computer, hence the bias. Unlikely to be the reason because: 1. the computer was not overloaded, 2. the same load would be seen by all benchmarked codes (that is precisely the purpose of "PerfTestUtils"). Anyways, this is all becoming rather moot since the previous figures are not reproducible... Best regards, Gilles - To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
Re: [rng] Usefulness of benchmarks
On Sun, Sep 4, 2016 at 2:06 PM, Gilleswrote: > A few quick runs seems to produce numbers that are now much > closer to what JMH produces.[1] > > What do you think? > How I can produce results with PerfTestUtils? Is there any script to run RandomStressTester? It possible that previous results were produced on loaded computer, hence the bias. Best regards, Artem Barger.
Re: [rng] Usefulness of benchmarks
Hi Artem. I've just updated the usage examples of Commons Math: commit ae8f5f04574be75727f5e948e04bc649cbdbbb3b A few quick runs seems to produce numbers that are now much closer to what JMH produces.[1] What do you think? I'm now fairly comfortable dropping the "PerfTestUtils" column from the table in the userguide (i.e. IMHO it's sufficiently close to what JMH provides to not raise concern). Thanks, Gilles [1] Although I don't recall having made any code update since the previous numbers were produced that could account for this change. :-{ On Sun, 04 Sep 2016 00:17:17 +0200, Gilles wrote: On Sun, 4 Sep 2016 00:33:32 +0300, Artem Barger wrote: On Sat, Sep 3, 2016 at 1:36 AM, Gilleswrote: The discrepancy between "PerfTestUtils" and JMH could be a bug (in "PerfTestUtils" of course!) or ... measuring different use-cases: Use of several RNGs at the same time vs using a single one; the latter could allow for more aggressive optimizations. I'm not really familiar with the PerfTestUtils, while I know that JMH is doing a great job to avoid different pitfalls while building microbenchmarks for measuring a performance. Also it looks a bit suspicious where comparing JDK random generator against itself it's not showing ration of 1.0 for PerfTestUtils. That is easy to explain: the ratio was done wrt to a "java.util.Random" object, while "RandomSource.JDK" wraps an instance of "java.util.Random". Lacking input as to what the benchmarks purport to demonstrate, I'm about to simply delete the "PerfTestUtils" column. The result will be a simplified (maybe simplistic) view of the relative performance of the RNGs in the "single use" use-case. I can try to take a look on PerfTestUtils to understand what is the main cause of such difference. AFAICS, the main difference is that JMH benchmarks one code (marked with "@Benchmark") after the other while "PerfTestUtils" benchmarks all the codes together. Any comment, objection, explanation, suggestion? [E.g. set up JMH to benchmark the other use case, or a reason why this is in fact not necessary.] We can play with different amount of warm up rounds in JMH to see whenever there is a degradation to results similar to PerfTestUtils for example. I'm pretty sure it's not related to warmup because the "PerfTestUtils" benchmarks were running for much longer than JMH. What would be interesting is to see whether JMH performance degrades when multiple RNGs (of different types) have been instantiated and run in the same scope. Regards, Gilles - To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
Re: [rng] Usefulness of benchmarks
On Sun, 4 Sep 2016 00:33:32 +0300, Artem Barger wrote: On Sat, Sep 3, 2016 at 1:36 AM, Gilleswrote: The discrepancy between "PerfTestUtils" and JMH could be a bug (in "PerfTestUtils" of course!) or ... measuring different use-cases: Use of several RNGs at the same time vs using a single one; the latter could allow for more aggressive optimizations. I'm not really familiar with the PerfTestUtils, while I know that JMH is doing a great job to avoid different pitfalls while building microbenchmarks for measuring a performance. Also it looks a bit suspicious where comparing JDK random generator against itself it's not showing ration of 1.0 for PerfTestUtils. That is easy to explain: the ratio was done wrt to a "java.util.Random" object, while "RandomSource.JDK" wraps an instance of "java.util.Random". Lacking input as to what the benchmarks purport to demonstrate, I'm about to simply delete the "PerfTestUtils" column. The result will be a simplified (maybe simplistic) view of the relative performance of the RNGs in the "single use" use-case. I can try to take a look on PerfTestUtils to understand what is the main cause of such difference. AFAICS, the main difference is that JMH benchmarks one code (marked with "@Benchmark") after the other while "PerfTestUtils" benchmarks all the codes together. Any comment, objection, explanation, suggestion? [E.g. set up JMH to benchmark the other use case, or a reason why this is in fact not necessary.] We can play with different amount of warm up rounds in JMH to see whenever there is a degradation to results similar to PerfTestUtils for example. I'm pretty sure it's not related to warmup because the "PerfTestUtils" benchmarks were running for much longer than JMH. What would be interesting is to see whether JMH performance degrades when multiple RNGs (of different types) have been instantiated and run in the same scope. Regards, Gilles - To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
Re: [rng] Usefulness of benchmarks
On Sat, Sep 3, 2016 at 1:36 AM, Gilleswrote: > The discrepancy between "PerfTestUtils" and JMH could be a bug (in > "PerfTestUtils" of course!) or ... measuring different use-cases: > Use of several RNGs at the same time vs using a single one; the > latter could allow for more aggressive optimizations. > I'm not really familiar with the PerfTestUtils, while I know that JMH is doing a great job to avoid different pitfalls while building microbenchmarks for measuring a performance. Also it looks a bit suspicious where comparing JDK random generator against itself it's not showing ration of 1.0 for PerfTestUtils. > Lacking input as to what the benchmarks purport to demonstrate, I'm > about to simply delete the "PerfTestUtils" column. > The result will be a simplified (maybe simplistic) view of the > relative performance of the RNGs in the "single use" use-case. > > I can try to take a look on PerfTestUtils to understand what is the main cause of such difference. > Any comment, objection, explanation, suggestion? > [E.g. set up JMH to benchmark the other use case, or a reason why > this is in fact not necessary.] > We can play with different amount of warm up rounds in JMH to see whenever there is a degradation to results similar to PerfTestUtils for example. Best regards, Artem Barger.