What does CompileThreshold, Tier2CompileThreshold, Tier3CompileThreshold and Tier4CompileThreshold control?

Question

HotSpot's tiered compilation uses the interpreter until a threshold of invocations (for methods) or iterations (for loops) triggers a client compilation with self-profiling. The client compilation is used until another threshold of invocations or iterations triggers a server compilation.

Printing HotSpot's flags shows the following flag values with -XX:+TieredCompilation.

intx CompileThreshold      = 10000 {pd product}         intx Tier2CompileThreshold = 0     {product}            intx Tier3CompileThreshold = 2000  {product}            intx Tier4CompileThreshold = 15000 {product}            

There are too many flags for just a client and server compiler. What compilers are controlled by these flags? If not client and server, what is the purpose of the additional compilers?

Are CompileThreshold and Tier2CompileThreshold ignored in this case? What does Tier3CompileThreshold control when a client compilation is triggered? What does Tier4CompileThreshold control when a server compilation is triggered?

Answer 1

The comments in advancedThresholdPolicy.hpp discuss the different compiler tiers and the thresholds. See that file for a deeper discussion.

The system supports 5 execution levels:

• Tier 0 - interpreter
• Tier 1 - C1 with full optimization (no profiling)
• Tier 2 - C1 with invocation and backedge counters
• Tier 3 - C1 with full profiling (level 2 + MDO)
• Tier 4 - C2

C1 is the client compiler. C2 is the server compiler.

In the common case, the compilation goes: 0 → 3 → 4. Atypical cases are used based on C1 and C2 queue lengths. Tier 2 is used when the C2 queue length is too long so that the method can execute about 30% faster until the C2 can process the profiling information. If the method is determined to be trivial, then it is compiled with Tier 1 since it will produce the same code as Tier 4.

Thresholds are dynamically adjusted based on the length of the C1 and C2 queues.