GCPerfP99V2.java - Object Number and Size

Java GC Tutorials - Herong's Tutorial Examples

∟Garbage Collection Performance Test Program

∟GCPerfP99V2.java - Object Number and Size

This section provides tutorials to find the work load level for the GCPerfP99V2.java program to the impact of execution environment interruptions to a very level. Work load of 1000 gives me a coefficient variation of 1.22% on my computer used in 2019.

In the last tutorial, I was able the increase the work load of my application to reduce the impact of environment interruptions to a very level.

In this tutorial, I want to increase the work load of the GC to increase its impact on my application, so that I can measure it more accurately.

Test 1 - Running GCPerfP99V2.java with the default object size and chunk size.

herong> java -XX:+UseSerialGC -Xms400m -Xmx400m -Xlog:gc \
   GCPerfP99V2 1024 32 32 0 10 1000 100

Parameters:
   Size=1024KB, Base=32, Chunk=32, Warmup=0, Runs=10, Loads=1000
[1.197s] Run start 1
[2.352s] Run end 1: 1155ms
[2.361s] Run start 2
[3.501s] Run end 2: 1149ms
[3.502s] Run start 3
GC(0) Pause Young (Allocation Failure) 106M->38M(386M) 18.735ms
[4.665s] Run end 3: 1164ms
[4.666s] Run start 4
[5.829s] Run end 4: 1164ms
[5.829s] Run start 5
[6.979s] Run end 5: 1150ms
[6.98s] Run start 6
GC(1) Pause Young (Allocation Failure) 144M->74M(386M) 22.835ms
[8.14s] Run end 6: 1161ms
[8.14s] Run start 7
[9.268s] Run end 7: 1128ms
[9.269s] Run start 8
[10.4s] Run end 8: 1132ms
[10.4s] Run start 9
GC(2) Pause Young (Allocation Failure) 180M->119M(386M) 23.174ms
[11.554s] Run end 9: 1154ms
[11.555s] Run start 10
[12.682s] Run end 10: 1128ms
Statistics of percentile runs:
runs = 10
percentile = 100
P99 = 10
   Total number of runs = 10
   Percentile % = 100
   Execution time = 11485 ms
   Average time per run = 1148 ms
   Standard deviation = 13.516656391282572 ms
   Objects processed = 320
   Throughput = 27 objects/second
   Latency = 37037 ms/1000 objects
   Throughput (max, ave, min) = (28, 27, 27)
   Latency (min, ave, max) = (35714, 37037, 37037)
Worst runs dropped:

As you can see, GC took 18 to 23 ms to complete, which is about the same level as other environment interruptions. So I need to increase the number of objects in the Eden for the GC to process. This can be done by decrease the object size and increase the chunk size to keep the same amount work for each run.

Test 2 - Running GCPerfP99V2.java with object size of 32K and chunk size of 1024.

herong> java -XX:+UseSerialGC -Xms400m -Xmx400m -Xlog:gc \
   GCPerfP99V2 32 1024 1024 0 10 1000 100

Parameters:
   Size=32KB, Base=1024, Chunk=1024, Warmup=0, Runs=10, Loads=1000
[1.204s] Run start 1
[2.349s] Run end 1: 1145ms
[2.358s] Run start 2
[3.496s] Run end 2: 1147ms
[3.497s] Run start 3
GC(0) Pause Young (Allocation Failure) 106M->38M(386M) 19.225ms
[4.651s] Run end 3: 1155ms
[4.651s] Run start 4
[5.785s] Run end 4: 1134ms
[5.786s] Run start 5
[6.919s] Run end 5: 1134ms
[6.92s] Run start 6
GC(1) Pause Young (Allocation Failure) 145M->138M(386M) 51.666ms
[8.102s] Run end 6: 1183ms
[8.103s] Run start 7
[9.232s] Run end 7: 1130ms
[9.232s] Run start 8
[10.365s] Run end 8: 1133ms
[10.366s] Run start 9
GC(2) Pause Young (Allocation Failure) 245M->244M(386M) 51.924ms
[11.55s] Run end 9: 1185ms
[11.551s] Run start 10
[12.685s] Run end 10: 1135ms
Statistics of percentile runs:
   Total number of runs = 10
   Percentile % = 100
   Execution time = 11481 ms
   Average time per run = 1148 ms
   Standard deviation = 19.39845354661036 ms
   Objects processed = 10240
   Throughput = 891 objects/second
   Latency = 1157 ms/1000 objects
   Throughput (max, ave, min) = (906, 891, 864)
   Latency (min, ave, max) = (1103, 1122, 1157)
Worst runs dropped: