Re: Annotation access speed comparison (sneak-peek, non-authoritative)
I also added selectCovered benchmarks with s.getBegin and s.getEnd. Weird, they seem to be even faster. GROUP: select covered by WARM-UP: Running benchmark... 10 100 1000 selectCovered(Token.class, s).forEach(t -> {}): Running benchmark... 10 100 1000 JCAS.select(Token.class).coveredBy(s).forEach(t -> {}): Running benchmark... 10 100 1000 JCAS.getAnnotationIndex(Token.class).select().coveredBy(s).forEach(t -> {}): Running benchmark... 10 100 1000 JCAS.select(Token.class).filter(t -> coveredBy(t, s)).forEach(t -> {}): Running benchmark... 10 100 1000 JCAS.getAnnotationIndex(Token.class).stream().filter(t -> coveredBy(t, s)).forEach(t -> {}): Running benchmark... 10 100 1000 JCAS.select(Token.class).coveredBy(s.getBegin(), s.getEnd()).forEach(t -> {}): Running benchmark... 10 100 1000 JCAS.getAnnotationIndex(Token.class).select().coveredBy(s.getBegin(), s.getEnd()).forEach(t -> {}): Running benchmark... 10 100 1000 selectCovered(JCAS, Token.class, s.getBegin(), s.getEnd()).forEach(t -> {}): Running benchmark... 10 100 1000 Sorted by execution time: 19308000ns / 7667000ns -- WARM-UP 97388000ns / 3862000ns -- JCAS.getAnnotationIndex(Token.class).select().coveredBy(s.getBegin(), s.getEnd()).forEach(t -> {}) 114012000ns / 7585000ns -- JCAS.getAnnotationIndex(Token.class).select().coveredBy(s).forEach(t -> {}) 139689000ns / 11349000ns -- JCAS.select(Token.class).coveredBy(s.getBegin(), s.getEnd()).forEach(t -> {}) 158952000ns / 30894000ns -- selectCovered(JCAS, Token.class, s.getBegin(), s.getEnd()).forEach(t -> {}) 189711000ns / 12981000ns -- selectCovered(Token.class, s).forEach(t -> {}) 233259000ns / 12051000ns -- JCAS.select(Token.class).coveredBy(s).forEach(t -> {}) 737271000ns / 30204000ns -- JCAS.getAnnotationIndex(Token.class).stream().filter(t -> coveredBy(t, s)).forEach(t -> {}) 846053000ns / 45272000ns -- JCAS.select(Token.class).filter(t -> coveredBy(t, s)).forEach(t -> {}) Cheers Mario > On 4 Nov 2020, at 16.03, Mario Juric wrote: > > Hi Richard, > > This is really useful, and it makes me much more positive about the select > API performance than the previous benchmarks suggested, but you need to know > what calling approach works best, e.g. getting the annotation index first and > doing select on this performs better. > > I added support for nano seconds CPU time to Benchmark and changed the > SelectBenchmark to use user timer, which I think is much more accurate about > the actual time spent in the select operations. However, in this case I don’t > think it really changes the overall conclusion to the normal system clock > version, although I haven’t compared the results line by line. You can find > the branch here: > > https://github.com/mjunsilo/uima-uimafit/tree/mjuric/benchmark-cpu-time > > Do I need to make a Jira issue for this? Let me first know if there is any > interest. > > Cheers > Mario > > > On 4 Nov 2020, at 11.08, Richard Eckart de Castilho > mailto:r...@apache.org>> wrote: > > External email – Do not click links or open attachments unless you recognize > the sender and know that the content is safe. > > > Hi all, > > for those who are interested - here are a few results of benchmarking the > access to annotations in the CAS using different approaches. > > These were done on a Macbook Pro (i7 2,2 GHz) basically under working > conditions (many applications open, etc.) > > Used versions: > - uimaj-core: 3.1.2-SNAPSHOT (commit 099a2e0a9) > - uimafit-core: 3.1.1-SNAPSHOT (commit 72895b5c8) > > The benchmarks basically fill a CAS with random annotations (Sentence and > Token type, but they do not behave > like sentences/tokens usually would - i.e. they are generated at random > positions and so may arbitrarily > overlap with each other). All annotations start/end within a range of [0-130] > and have a random length between > 0 and 30. The benchmarks fill the CAS multiple times with increasing numbers > of annotations and perform the > selections repeatedly. If you want more details, check out the > uimafit-benchmark module and run them yourself ;) > The first timing is the cumulative time spend by benchmark. The second timing > is the longest duration of a > single execution. > > > As for insights: > > * Don't look at the times in terms of absolute values - rather consider how > the time of one approach > behaves relative to the time of another approach. > > * I find it quite interesting that selects are slower when using > JCAS.select(type) than when using > JCAS.getAnnotationIndex(type).select(). I would expect both to run at the > same speed. > > * Contrary to previous benchmark results, we can see that the (J)CAS.select() > is typically faster than > its uimaFIT counterpart with a few interesting exceptions. > > * Note that there is no CAS.select().overlapping() equivalent to the > JCasUtil.selectOverlapping (yet) > > > If you would like to see additional approaches measured or if you have ideas > of how to improve the >
Re: Annotation access speed comparison (sneak-peek, non-authoritative)
Hi Richard, This is really useful, and it makes me much more positive about the select API performance than the previous benchmarks suggested, but you need to know what calling approach works best, e.g. getting the annotation index first and doing select on this performs better. I added support for nano seconds CPU time to Benchmark and changed the SelectBenchmark to use user timer, which I think is much more accurate about the actual time spent in the select operations. However, in this case I don’t think it really changes the overall conclusion to the normal system clock version, although I haven’t compared the results line by line. You can find the branch here: https://github.com/mjunsilo/uima-uimafit/tree/mjuric/benchmark-cpu-time Do I need to make a Jira issue for this? Let me first know if there is any interest. Cheers Mario On 4 Nov 2020, at 11.08, Richard Eckart de Castilho mailto:r...@apache.org>> wrote: External email – Do not click links or open attachments unless you recognize the sender and know that the content is safe. Hi all, for those who are interested - here are a few results of benchmarking the access to annotations in the CAS using different approaches. These were done on a Macbook Pro (i7 2,2 GHz) basically under working conditions (many applications open, etc.) Used versions: - uimaj-core: 3.1.2-SNAPSHOT (commit 099a2e0a9) - uimafit-core: 3.1.1-SNAPSHOT (commit 72895b5c8) The benchmarks basically fill a CAS with random annotations (Sentence and Token type, but they do not behave like sentences/tokens usually would - i.e. they are generated at random positions and so may arbitrarily overlap with each other). All annotations start/end within a range of [0-130] and have a random length between 0 and 30. The benchmarks fill the CAS multiple times with increasing numbers of annotations and perform the selections repeatedly. If you want more details, check out the uimafit-benchmark module and run them yourself ;) The first timing is the cumulative time spend by benchmark. The second timing is the longest duration of a single execution. As for insights: * Don't look at the times in terms of absolute values - rather consider how the time of one approach behaves relative to the time of another approach. * I find it quite interesting that selects are slower when using JCAS.select(type) than when using JCAS.getAnnotationIndex(type).select(). I would expect both to run at the same speed. * Contrary to previous benchmark results, we can see that the (J)CAS.select() is typically faster than its uimaFIT counterpart with a few interesting exceptions. * Note that there is no CAS.select().overlapping() equivalent to the JCasUtil.selectOverlapping (yet) If you would like to see additional approaches measured or if you have ideas of how to improve the informativeness or general setup of the benchmarks, let me know. For small changes, you could also just open a PR on GitHub against uimaFIT master. Cheers, -- Richard GROUP: select = Sorted by execution time: 1136ms /2ms -- JCAS.select(Token.class).forEach(x -> {}) 1231ms /3ms -- JCasUtil.select(JCAS, Token.class).forEach(x -> {}) 2679ms /4ms -- JCAS.select(TOP.class).forEach(x -> {}) 2703ms /4ms -- JCAS.select().forEach(x -> {}) 3803ms /6ms -- JCasUtil.select(JCAS, TOP.class).forEach(x -> {}) 3997ms / 16ms -- JCasUtil.selectAll(JCAS).forEach(x -> {}) GROUP: select covered by = Sorted by execution time: 84ms /5ms -- JCAS.getAnnotationIndex(Token.class).select().coveredBy(s).forEach(t -> {}) 134ms / 11ms -- JCasUtil.selectCovered(Token.class, s).forEach(t -> {}) 159ms / 11ms -- JCAS.select(Token.class).coveredBy(s).forEach(t -> {}) 836ms / 46ms -- JCAS.getAnnotationIndex(Token.class).stream().filter(t -> coveredBy(t, s)).forEach(t -> {}) 842ms / 46ms -- JCAS.select(Token.class).filter(t -> coveredBy(t, s)).forEach(t -> {}) GROUP: select covering = Sorted by execution time: 98ms /5ms -- JCAS.getAnnotationIndex(Token.class).select().covering(s).forEach(t -> {}) 109ms /6ms -- CAS.getAnnotationIndex(getType(cas, TYPE_NAME_TOKEN)).select().covering(s).forEach(t -> {}) 157ms /7ms -- CasUtil.selectCovering(tokenType, s).forEach(t -> {}) 170ms / 20ms -- JCasUtil.selectCovering(Token.class, s).forEach(t -> {}) 187ms / 14ms -- JCAS.select(Token.class).covering(s).forEach(t -> {}) 812ms / 47ms -- JCAS.select(Token.class).filter(t -> covering(t, s)).forEach(t -> {}) 862ms / 45ms -- CAS.getAnnotationIndex(getType(cas, TYPE_NAME_TOKEN)).stream().filter(t -> covering(t, s)).forEach(t -> {}) 1039ms / 65ms -- JCAS.getAnnotationIndex(Token.class).stream().filter(t -> covering(t, s)).forEach(t -> {}) GROUP: select at = Sorted by execution time: 31ms /2ms -- JCAS.select(Token.class).at(s).forEach(t -> {}) 65ms /
Re: UIMA-AS with ActiveMQ VM Transport
The vm is currently reserved for async aggregate AE (collocated) endpoints. Async aggregate AE communicates with its colocated delegate AE using java (non-jms) queues to avoid serialization/deserialization of CASes. The vm keyword in the endpoint is used to identify such collocated AEs. Don't remember why vm was chosen to represent these endpoints. This decision was made early in the development of UIMA-AS. Fixing that will require refactoring of the UIMA-AS aggregate code and possibly other components as well. Even if this is fixed, the UIMA-AS client still has to seriallize/deserialize CASes when messaging the service. Not sure if using the vm protocol would reduce the latency assuming this is the main reason for using the AMQ vm transport between the UIMA-AS client and the service. -jerry On Tue, Nov 3, 2020 at 3:11 PM Richard Eckart de Castilho wrote: > Hi Jerry, > > > On 3. Nov 2020, at 17:52, Jaroslaw Cwiklik wrote: > > > > My apologies for a late reply to your question. The UIMA-AS client does > not > > support vm protocol between clients and services. Only tcp and http are > > currently supported as you've noted. You can manage broker persistence > > settings in the broker configuration See > > http://activemq.apache.org/persistence ( > > ) > > is there a particular (technical) reason why the vm protocol is not > supported? > > Would a potential interested contributor be faced with significant trouble > in > trying to add support for it? > > Cheers, > > -- Richard
Annotation access speed comparison (sneak-peek, non-authoritative)
Hi all, for those who are interested - here are a few results of benchmarking the access to annotations in the CAS using different approaches. These were done on a Macbook Pro (i7 2,2 GHz) basically under working conditions (many applications open, etc.) Used versions: - uimaj-core: 3.1.2-SNAPSHOT (commit 099a2e0a9) - uimafit-core: 3.1.1-SNAPSHOT (commit 72895b5c8) The benchmarks basically fill a CAS with random annotations (Sentence and Token type, but they do not behave like sentences/tokens usually would - i.e. they are generated at random positions and so may arbitrarily overlap with each other). All annotations start/end within a range of [0-130] and have a random length between 0 and 30. The benchmarks fill the CAS multiple times with increasing numbers of annotations and perform the selections repeatedly. If you want more details, check out the uimafit-benchmark module and run them yourself ;) The first timing is the cumulative time spend by benchmark. The second timing is the longest duration of a single execution. As for insights: * Don't look at the times in terms of absolute values - rather consider how the time of one approach behaves relative to the time of another approach. * I find it quite interesting that selects are slower when using JCAS.select(type) than when using JCAS.getAnnotationIndex(type).select(). I would expect both to run at the same speed. * Contrary to previous benchmark results, we can see that the (J)CAS.select() is typically faster than its uimaFIT counterpart with a few interesting exceptions. * Note that there is no CAS.select().overlapping() equivalent to the JCasUtil.selectOverlapping (yet) If you would like to see additional approaches measured or if you have ideas of how to improve the informativeness or general setup of the benchmarks, let me know. For small changes, you could also just open a PR on GitHub against uimaFIT master. Cheers, -- Richard GROUP: select = Sorted by execution time: 1136ms /2ms -- JCAS.select(Token.class).forEach(x -> {}) 1231ms /3ms -- JCasUtil.select(JCAS, Token.class).forEach(x -> {}) 2679ms /4ms -- JCAS.select(TOP.class).forEach(x -> {}) 2703ms /4ms -- JCAS.select().forEach(x -> {}) 3803ms /6ms -- JCasUtil.select(JCAS, TOP.class).forEach(x -> {}) 3997ms / 16ms -- JCasUtil.selectAll(JCAS).forEach(x -> {}) GROUP: select covered by = Sorted by execution time: 84ms /5ms -- JCAS.getAnnotationIndex(Token.class).select().coveredBy(s).forEach(t -> {}) 134ms / 11ms -- JCasUtil.selectCovered(Token.class, s).forEach(t -> {}) 159ms / 11ms -- JCAS.select(Token.class).coveredBy(s).forEach(t -> {}) 836ms / 46ms -- JCAS.getAnnotationIndex(Token.class).stream().filter(t -> coveredBy(t, s)).forEach(t -> {}) 842ms / 46ms -- JCAS.select(Token.class).filter(t -> coveredBy(t, s)).forEach(t -> {}) GROUP: select covering = Sorted by execution time: 98ms /5ms -- JCAS.getAnnotationIndex(Token.class).select().covering(s).forEach(t -> {}) 109ms /6ms -- CAS.getAnnotationIndex(getType(cas, TYPE_NAME_TOKEN)).select().covering(s).forEach(t -> {}) 157ms /7ms -- CasUtil.selectCovering(tokenType, s).forEach(t -> {}) 170ms / 20ms -- JCasUtil.selectCovering(Token.class, s).forEach(t -> {}) 187ms / 14ms -- JCAS.select(Token.class).covering(s).forEach(t -> {}) 812ms / 47ms -- JCAS.select(Token.class).filter(t -> covering(t, s)).forEach(t -> {}) 862ms / 45ms -- CAS.getAnnotationIndex(getType(cas, TYPE_NAME_TOKEN)).stream().filter(t -> covering(t, s)).forEach(t -> {}) 1039ms / 65ms -- JCAS.getAnnotationIndex(Token.class).stream().filter(t -> covering(t, s)).forEach(t -> {}) GROUP: select at = Sorted by execution time: 31ms /2ms -- JCAS.select(Token.class).at(s).forEach(t -> {}) 65ms /4ms -- JCAS.select(Token.class).at(s.getBegin(), s.getEnd()).forEach(t -> {}) 109ms / 29ms -- JCasUtil.selectAt(CAS, Token.class, s.getBegin(), s.getEnd()).forEach(t -> {}) 880ms / 41ms -- JCAS.getAnnotationIndex(Token.class).stream().filter(t -> colocated(t, s)).forEach(t -> {}) 936ms / 47ms -- JCAS.select(Token.class).filter(t -> colocated(t, s)).forEach(t -> {}) GROUP: select overlapping = Sorted by execution time: 238ms / 34ms -- JCasUtil.selectOverlapping(JCAS, Token.class, s).forEach(t -> {}) 354ms / 22ms -- JCAS.getAnnotationIndex(Token.class).stream().filter(t -> overlapping(t, s)).forEach(t -> {}) 381ms / 24ms -- CAS.select(Token.class).filter(t -> overlapping(t, s)).forEach(t -> {})