Using the Falcon Compiler
The Falcon compiler is the default optimizing JIT (Just-In-Time) compiler in the Azul Zulu Prime Builds of OpenJDK (Azul Zulu Prime JVM).
Falcon Optimization Notes
The main focus of the Falcon compiler is generating high performance code. This may result in longer compile times because additional CPU time is required for higher levels of optimization. Longer Falcon compile times can increase the application startup and warm-up times when compared to the startup and warm-up times when running Azul Zulu Prime JVM using the C2 compiler.
Falcon fully supports Azul ReadyNow! technology, which helps improve consistency of application performance during warmup.
While future versions of Azul Zulu Prime JVM will likely include automatic ongoing eviction of contents in the Falcon cache, the current Azul Zulu Prime JVM version does not, and as of version 18.01.0.0, the Falcon cache is not automatically cleared or evicted. The cache will accumulate contents up to
FalconMaxCacheSize. Once it reaches that size, newly encountered compilations will no longer be added to the cache. The cache can be reset at any time (including while the Azul Zulu Prime JVM is running) by destroying the contents of the cache with, for example,
rm -r .zing-falcon-cache.
History and Alternatives
The Falcon compiler first became available as a feature in Azul Platform Prime 126.96.36.199. Falcon became the default Tier 2 optimizing compiler in Azul Platform Prime 17.03.0.0. If, for some reason, you need to switch to the non-default Tier 2 C2 compiler in Azul Platform Prime 17.03.0.0 - 20.03.1.0, use the
-XX:+UseC2 option. For example,
$JAVA_HOME/bin/java -XX:+UseC2 –cp . myprog
Starting with Azul Platform Prime 20.04.0.0, the
-XX:+UseC2 option can use one of two separate implementations of C2 JIT compilation:
A traditional SeaOfNodesC2 mode. A C2 implementation that has been part of Azul Platform Prime since 2010.
A new KestrelC2 mode. A newer C2 implementation introduced to Azul Platform Prime in 2020. It is based on a lightweight use of the LLVM backend, and typically produces faster code than SeaOfNodesC2 while keeping compilation effort at similar levels.
Both KestrelC2 and SeaOfNodesC2 typically exhibit significantly lower compilation-time CPU consumption than Falcon.
Note that while C2 is entirely supported, active development has mostly stopped. As such, switching to C2 should be seen as a workaround until a Falcon issue can be resolved, and you are encouraged to retry Falcon after every major upgrade.
Run Zing Vision Robot (ZVRobot), the profiling and diagnostic snapshot collection tool for the Azul Zulu Prime JVM, together with the application. Basic ZVRobot Usage:
To the Azul Zulu Prime JVM command-line options add:
Optionally, configure the VM to collect compiler intermediate representation (IR) dumps for sharing with Azul Support
-XX:FalconIRDumpRoot=<path>– sets the directory in which to dump IR. If not set, defaults to
-XX:FalconDumpIRToDiskOf=*– tells the Falcon compiler to dump its internal IR for compiles which match the specified pattern to disk. This IR can be provided to Azul Support to assist in analyzing performance issues. Enabling this option may affect the performance of the application. This option should not be used outside of a test environment.
Configure ZVRobot to connect to the Azul Zulu Prime JVM by modifying the ZVRobot properties file. You can configure ZVRobot to collect different amounts of information.
Start ZVRobot using the modified properties file created in step 3. Example command:
java -jar ZVRobot-<version>.jar <output_directory> ZVRobot.prop.
Cloud Native Compiler
Azul’s Cloud Native Compiler can further improve the performance of the Falcon compiler by removing JIT compilation from your JVM and running them on a dedicate server. By offloading JIT compilation to the server, you get faster warm-up times with lower CPU consumption on your client JVMs. See the Cloud Native Compiler documentation for more information.