[
https://issues.apache.org/jira/browse/IGNITE-23049?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Alexander Lapin updated IGNITE-23049:
-------------------------------------
Description:
h3. Motivation
Originally HybridClockImpl used
{code:java}
systemUTC().instant().toEpochMilli(){code}
within Reduce latency for txn preparing phase it was substituted with
{code:java}
FastTimestamps.coarseCurrentTimeMillis(){code}
that likely led to decreased test stability:
#
[ClockWaiterTest.futureCompletesWithoutClockUpdatesfailure|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8409117?hideProblemsFromDependencies=false&hideTestsFromDependencies=false&expandBuildDeploymentsSection=false&expandCode+Inspection=true&expandBuildProblemsSection=true&expandBuildChangesSection=true&expandBuildTestsSection=true]
#
[IndexMetaStorageRecoveryTest.testMissingDropTable|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8411374?hideProblemsFromDependencies=false&hideTestsFromDependencies=false&expandCode+Inspection=true&expandBuildChangesSection=true&expandBuildDeploymentsSection=false&expandBuildProblemsSection=true&expandBuildTestsSection=true]
#
[ItIgniteDistributionZoneManagerNodeRestartTest.testScaleDownTimerIsRestoredAfterRestart(defaultZone=true)|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8414582?hideTestsFromDependencies=false&hideProblemsFromDependencies=false&expandBuildTestsSection=true&expandBuildDeploymentsSection=false&expandCode+Inspection=true&expandBuildProblemsSection=true&expandBuildChangesSection=true]
Thus, it seems reasonable to benchmark following basis for HybridClock in order
to choose the fastest and most reliable
* systemUTC().instant().toEpochMilli()
* System.currentTimeMillis()
* FastTimestamps.coarseCurrentTimeMillis()
h3. Defition of Done
* Write benchmarks and compare results.
h3. UPD№1
In average, and what's more important according to UpsertKvBenchmark
{*}FastTimestamps ≈ currentTimeMillis < systemUTC{*}, ** thus we may substitute
unstable FastsTimestamps with currentTimeMillis after proving its stability.
Benchmark results for three HybridClock implementations within 1, 5 and 10
threads:
{code:java}
Chip: Apple M3 Pro Total Number of Cores: 11 (5 performance and 6 efficiency)
Memory: 36 GB
Benchmark (implementationName)
Mode Cnt Score Error Units
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockSystemUTC
avgt 10 515.640 ± 39.361 us/op
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockCurrentTimeMillis
avgt 10 253.520 ± 39.885 us/op
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockFastTimestamps
avgt 10 218.364 ± 14.804 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockSystemUTC
avgt 10 48.886 ± 3.840 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockCurrentTimeMillis
avgt 10 14.379 ± 0.074 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockFastTimestamps
avgt 10 4.153 ± 0.199 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockSystemUTC
avgt 10 1498.397 ± 19.352 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockCurrentTimeMillis
avgt 10 847.807 ± 8.346 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockFastTimestamps
avgt 10 3027.382 ± 175.118 us/op
Win 10, 2 cores
Benchmark (implementationName)
Mode Cnt Score Error Units
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockSystemUTC
avgt 10 376.678 � 5.800 us/op
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockCurrentTimeMillis
avgt 10 329.343 � 4.630 us/op
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockFastTimestamps
avgt 10 259.633 � 7.708 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockSystemUTC
avgt 10 24.561 � 0.285 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockCurrentTimeMillis
avgt 10 9.907 � 0.061 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockFastTimestamps
avgt 10 10.102 � 0.186 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockSystemUTC
avgt 10 672.293 � 11.405 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockCurrentTimeMillis
avgt 10 534.624 � 31.084 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockFastTimestamps
avgt 10 526.803 � 14.244 us/op
Debian 12, CPU=intel 8250u 4 cores, 48GB. OpenJDK 11.0.20
Benchmark (implementationName)
Mode Cnt Score Error Units
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockSystemUTC
avgt 10 671.298 ± 20.599 us/op
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockCurrentTimeMillis
avgt 10 780.638 ± 62.643 us/op
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockFastTimestamps
avgt 10 1182.727 ± 960.559 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockSystemUTC
avgt 10 46.180 ± 0.986 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockCurrentTimeMillis
avgt 10 29.254 ± 0.503 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockFastTimestamps
avgt 10 9.760 ± 0.127 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockSystemUTC
avgt 10 1509.756 ± 60.812 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockCurrentTimeMillis
avgt 10 1407.827 ± 22.288 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockFastTimestamps
avgt 10 1835.972 ± 172.746 us/op
mac os m1, 10 cores
Benchmark (implementationName)
Mode Cnt Score Error Units
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockSystemUTC
avgt 10 614.132 ± 331.664 us/op
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockCurrentTimeMillis
avgt 10 330.902 ± 13.540 us/op
HybridClockBenchmark.hybridClockNowFiveThreads HybridClockFastTimestamps
avgt 10 156.549 ± 10.527 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockSystemUTC
avgt 10 87.725 ± 1.370 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockCurrentTimeMillis
avgt 10 21.972 ± 3.774 us/op
HybridClockBenchmark.hybridClockNowSingleThread HybridClockFastTimestamps
avgt 10 8.082 ± 0.341 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockSystemUTC
avgt 10 1166.027 ± 50.319 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockCurrentTimeMillis
avgt 10 646.750 ± 58.910 us/op
HybridClockBenchmark.hybridClockNowTenThreads HybridClockFastTimestamps
avgt 10 297.513 ± 22.518 us/op {code}
and single-threaded UpsertKvBenchmark
{code:java}
Chip: Apple M3 Pro Total Number of Cores: 11 (5 performance and 6 efficiency)
Memory: 36 GB
systemUTC
Benchmark (fsync) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert false avgt 10 9.150 ± 0.265 us/op
UpsertKvBenchmark.upsert true avgt 10 6.401 ± 0.349 us/op
currentTimeMillis
Benchmark (fsync) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert false avgt 10 6.535 ± 0.533 us/op
UpsertKvBenchmark.upsert true avgt 10 5.282 ± 0.265 us/op
FastTimestamps
Benchmark (fsync) Mode Cnt Score Error Units
UpsertKvBenchmark.upsert false avgt 10 6.957 ± 0.449 us/op
UpsertKvBenchmark.upsert true avgt 10 5.237 ± 0.260 us/op {code}
was:
h3. Motivation
Originally HybridClockImpl used
{code:java}
systemUTC().instant().toEpochMilli(){code}
within Reduce latency for txn preparing phase it was substituted with
{code:java}
FastTimestamps.coarseCurrentTimeMillis(){code}
that likely led to decreased test stability:
#
[ClockWaiterTest.futureCompletesWithoutClockUpdatesfailure|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8409117?hideProblemsFromDependencies=false&hideTestsFromDependencies=false&expandBuildDeploymentsSection=false&expandCode+Inspection=true&expandBuildProblemsSection=true&expandBuildChangesSection=true&expandBuildTestsSection=true]
#
[IndexMetaStorageRecoveryTest.testMissingDropTable|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8411374?hideProblemsFromDependencies=false&hideTestsFromDependencies=false&expandCode+Inspection=true&expandBuildChangesSection=true&expandBuildDeploymentsSection=false&expandBuildProblemsSection=true&expandBuildTestsSection=true]
#
[ItIgniteDistributionZoneManagerNodeRestartTest.testScaleDownTimerIsRestoredAfterRestart(defaultZone=true)|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8414582?hideTestsFromDependencies=false&hideProblemsFromDependencies=false&expandBuildTestsSection=true&expandBuildDeploymentsSection=false&expandCode+Inspection=true&expandBuildProblemsSection=true&expandBuildChangesSection=true]
Thus, it seems reasonable to benchmark following basis for HybridClock in order
to choose the fastest and most reliable
* systemUTC().instant().toEpochMilli()
* System.currentTimeMillis()
* FastTimestamps.coarseCurrentTimeMillis()
h3. Defition of Done
* Write benchmarks and compare results.
> Benchmark HybridClock
> ---------------------
>
> Key: IGNITE-23049
> URL: https://issues.apache.org/jira/browse/IGNITE-23049
> Project: Ignite
> Issue Type: Improvement
> Reporter: Alexander Lapin
> Assignee: Alexander Lapin
> Priority: Major
> Labels: ignite-3
> Time Spent: 10m
> Remaining Estimate: 0h
>
> h3. Motivation
> Originally HybridClockImpl used
> {code:java}
> systemUTC().instant().toEpochMilli(){code}
> within Reduce latency for txn preparing phase it was substituted with
> {code:java}
> FastTimestamps.coarseCurrentTimeMillis(){code}
> that likely led to decreased test stability:
> #
> [ClockWaiterTest.futureCompletesWithoutClockUpdatesfailure|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8409117?hideProblemsFromDependencies=false&hideTestsFromDependencies=false&expandBuildDeploymentsSection=false&expandCode+Inspection=true&expandBuildProblemsSection=true&expandBuildChangesSection=true&expandBuildTestsSection=true]
> #
> [IndexMetaStorageRecoveryTest.testMissingDropTable|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8411374?hideProblemsFromDependencies=false&hideTestsFromDependencies=false&expandCode+Inspection=true&expandBuildChangesSection=true&expandBuildDeploymentsSection=false&expandBuildProblemsSection=true&expandBuildTestsSection=true]
> #
> [ItIgniteDistributionZoneManagerNodeRestartTest.testScaleDownTimerIsRestoredAfterRestart(defaultZone=true)|https://ci.ignite.apache.org/buildConfiguration/ApacheIgnite3xGradle_Test_RunAllTests/8414582?hideTestsFromDependencies=false&hideProblemsFromDependencies=false&expandBuildTestsSection=true&expandBuildDeploymentsSection=false&expandCode+Inspection=true&expandBuildProblemsSection=true&expandBuildChangesSection=true]
> Thus, it seems reasonable to benchmark following basis for HybridClock in
> order to choose the fastest and most reliable
> * systemUTC().instant().toEpochMilli()
> * System.currentTimeMillis()
> * FastTimestamps.coarseCurrentTimeMillis()
> h3. Defition of Done
> * Write benchmarks and compare results.
> h3. UPD№1
> In average, and what's more important according to UpsertKvBenchmark
> {*}FastTimestamps ≈ currentTimeMillis < systemUTC{*}, ** thus we may
> substitute unstable FastsTimestamps with currentTimeMillis after proving its
> stability.
> Benchmark results for three HybridClock implementations within 1, 5 and 10
> threads:
>
> {code:java}
> Chip: Apple M3 Pro Total Number of Cores: 11 (5 performance and 6 efficiency)
> Memory: 36 GB
> Benchmark (implementationName)
> Mode Cnt Score Error Units
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockSystemUTC
> avgt 10 515.640 ± 39.361 us/op
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockCurrentTimeMillis
> avgt 10 253.520 ± 39.885 us/op
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockFastTimestamps
> avgt 10 218.364 ± 14.804 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockSystemUTC
> avgt 10 48.886 ± 3.840 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockCurrentTimeMillis
> avgt 10 14.379 ± 0.074 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockFastTimestamps
> avgt 10 4.153 ± 0.199 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockSystemUTC
> avgt 10 1498.397 ± 19.352 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockCurrentTimeMillis
> avgt 10 847.807 ± 8.346 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockFastTimestamps
> avgt 10 3027.382 ± 175.118 us/op
> Win 10, 2 cores
> Benchmark (implementationName)
> Mode Cnt Score Error Units
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockSystemUTC
> avgt 10 376.678 � 5.800 us/op
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockCurrentTimeMillis
> avgt 10 329.343 � 4.630 us/op
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockFastTimestamps
> avgt 10 259.633 � 7.708 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockSystemUTC
> avgt 10 24.561 � 0.285 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockCurrentTimeMillis
> avgt 10 9.907 � 0.061 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockFastTimestamps
> avgt 10 10.102 � 0.186 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockSystemUTC
> avgt 10 672.293 � 11.405 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockCurrentTimeMillis
> avgt 10 534.624 � 31.084 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockFastTimestamps
> avgt 10 526.803 � 14.244 us/op
> Debian 12, CPU=intel 8250u 4 cores, 48GB. OpenJDK 11.0.20
> Benchmark (implementationName)
> Mode Cnt Score Error Units
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockSystemUTC
> avgt 10 671.298 ± 20.599 us/op
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockCurrentTimeMillis
> avgt 10 780.638 ± 62.643 us/op
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockFastTimestamps
> avgt 10 1182.727 ± 960.559 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockSystemUTC
> avgt 10 46.180 ± 0.986 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockCurrentTimeMillis
> avgt 10 29.254 ± 0.503 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockFastTimestamps
> avgt 10 9.760 ± 0.127 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockSystemUTC
> avgt 10 1509.756 ± 60.812 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockCurrentTimeMillis
> avgt 10 1407.827 ± 22.288 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockFastTimestamps
> avgt 10 1835.972 ± 172.746 us/op
> mac os m1, 10 cores
> Benchmark (implementationName)
> Mode Cnt Score Error Units
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockSystemUTC
> avgt 10 614.132 ± 331.664 us/op
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockCurrentTimeMillis
> avgt 10 330.902 ± 13.540 us/op
> HybridClockBenchmark.hybridClockNowFiveThreads HybridClockFastTimestamps
> avgt 10 156.549 ± 10.527 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockSystemUTC
> avgt 10 87.725 ± 1.370 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockCurrentTimeMillis
> avgt 10 21.972 ± 3.774 us/op
> HybridClockBenchmark.hybridClockNowSingleThread HybridClockFastTimestamps
> avgt 10 8.082 ± 0.341 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockSystemUTC
> avgt 10 1166.027 ± 50.319 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockCurrentTimeMillis
> avgt 10 646.750 ± 58.910 us/op
> HybridClockBenchmark.hybridClockNowTenThreads HybridClockFastTimestamps
> avgt 10 297.513 ± 22.518 us/op {code}
> and single-threaded UpsertKvBenchmark
>
>
> {code:java}
> Chip: Apple M3 Pro Total Number of Cores: 11 (5 performance and 6 efficiency)
> Memory: 36 GB
> systemUTC
> Benchmark (fsync) Mode Cnt Score Error Units
> UpsertKvBenchmark.upsert false avgt 10 9.150 ± 0.265 us/op
> UpsertKvBenchmark.upsert true avgt 10 6.401 ± 0.349 us/op
> currentTimeMillis
> Benchmark (fsync) Mode Cnt Score Error Units
> UpsertKvBenchmark.upsert false avgt 10 6.535 ± 0.533 us/op
> UpsertKvBenchmark.upsert true avgt 10 5.282 ± 0.265 us/op
> FastTimestamps
> Benchmark (fsync) Mode Cnt Score Error Units
> UpsertKvBenchmark.upsert false avgt 10 6.957 ± 0.449 us/op
> UpsertKvBenchmark.upsert true avgt 10 5.237 ± 0.260 us/op {code}
>
>
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
