{"id":232717,"date":"2025-12-16T11:17:27","date_gmt":"2025-12-16T19:17:27","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/java\/?p=232717"},"modified":"2025-12-16T15:09:50","modified_gmt":"2025-12-16T23:09:50","slug":"beyond-ergonomics-how-the-azure-command-launcher-for-java-improves-gc-stability-and-throughput-on-azure-vms","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/java\/beyond-ergonomics-how-the-azure-command-launcher-for-java-improves-gc-stability-and-throughput-on-azure-vms\/","title":{"rendered":"Beyond Ergonomics: How the Azure Command Launcher for Java Improves GC Stability and Throughput on Azure VMs"},"content":{"rendered":"

In our previous blog<\/a> we introduced Azure Command Launcher for Java (jaz<\/code>) \u2014a safe, resource-aware way to launch the JVM without hand-tuning dozens of flags. This follow-up shares performance results, focusing on how jaz<\/code> affects G1 behavior, heap dynamics, and pause characteristics under a long-running, allocation-intensive workload: SPECjbb 2015 (JBB).<\/p>\n

Test bed<\/strong>: 4-vCPU, 16-GB Azure Linux\/Arm64 VM running the Microsoft Build of OpenJDK.<\/p>\n

JDKs exercised<\/strong>: Validated on JDK 17 (17.0.17), 21 (21.0.9), and 25 (25.0.1); all figures in this post are from the JDK 17 runs. Trends on 21\/25 matched the 17 results.<\/p>\n

How we ran it:<\/strong><\/p>\n

# baseline\r\njava -jar specjbb.jar\r\n\r\n# with jaz\r\njaz -jar specjbb.jar<\/code><\/pre>\n
Controls<\/strong>: Same JBB workload config, OS settings, and JVM flags for both runs\u2014the launcher was the only change.<\/p>\n
SPECjbb 2015 (JBB) is a SPEC benchmark; we report relative trends only and do not publish raw scores.<\/div><\/p>\n
Why JBB Is the Right Stress Test<\/h2>\n
JBB exercises high allocation rate, object churn, humongous-allocation behavior, generational sizing, region pressure, concurrent-mark sustainability, GC scheduling, and pause-time predictability. Because it is both bandwidth-intensive and latency-sensitive, JBB is ideal for validating heap ergonomics and GC policies in the cloud.<\/p>\n
As a capacity-planning tool it helps explore sustained throughput limits, GC headroom before SLA violations, warm-up behavior under load, and how a given VM size (4 cores, 16 GB) holds up under continuous allocation pressure.<\/p>\n
\nFor detailed methodology and the JBB phase guide, see Appendix A<\/a> and Appendix B<\/a>.\nA GC refresher and figure legend are in Appendix C<\/a> and Appendix D<\/a>.<\/div><\/p>\n
Performance Summary: Baseline vs jaz<\/code><\/h2>\n\n\n\n\n\n\n\n\n\n\n\n
Metric<\/strong><\/td>\nBaseline<\/strong><\/td>\nWith jaz<\/code><\/strong><\/td>\nImprovement<\/strong><\/td>\n<\/tr>\n
Peak Throughput<\/strong><\/td>\nBaseline<\/td>\n22%<\/td>\nHigher max-jOPS<\/td>\n<\/tr>\n
SLA Performance<\/strong><\/td>\nBaseline<\/td>\n15%<\/td>\nHigher critical-jOPS<\/td>\n<\/tr>\n
Total GC Events<\/strong><\/td>\n3777<\/td>\n2526<\/td>\n-33% (1251 fewer)<\/td>\n<\/tr>\n
Young GC Count<\/strong><\/td>\n1100<\/td>\n1596<\/td>\n+45% (handles higher load)<\/td>\n<\/tr>\n
Mixed GC Count<\/strong><\/td>\n778<\/td>\n265<\/td>\n-66% (513 fewer)<\/td>\n<\/tr>\n
Young GC Overhead<\/strong><\/td>\n1.41%<\/td>\n2.60%<\/td>\nHigher but efficient<\/td>\n<\/tr>\n
Mixed GC Overhead<\/strong><\/td>\n0.96<\/td>\n0.39<\/td>\n-59% reduction<\/td>\n<\/tr>\n
Old Gen Pattern<\/strong><\/td>\nFlat plateau (600-900)<\/td>\nDeep sawtooth (200-1000)<\/td>\nDynamic sizing active<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
Key Insight:<\/strong> jaz<\/code> achieves 22% higher throughput by keeping Young GC efficient\u2014objects die in Eden\/Survivor instead of promoting prematurely to Old gen, dramatically reducing expensive Mixed GC work.<\/p>\n
Baseline Behavior: Where the Wild Things Are<\/h2>\n
Microsoft Build of OpenJDK with default G1 GC ergonomics. Long JBB run on a 4-core VM<\/em>.<\/p>\n
Region Dynamics: Tight Band with Sustained Old Gen Plateau<\/h3>\n