Code, measurements and tools for my blog series "Taking JVM on a performance ride (4D tensor)" at mmpataki.github.io.
The series takes one small benchmark, reading a million elements from a 4D float tensor (64 x 16 x 28 x 32), and follows what the JIT compiler, the CPU and the operating system do with it.
| Path | Contents |
|---|---|
src/main/java/com/mpataki |
Tensor4D and the JMH benchmark Tensor4DBenchmark |
observations/p1-baseline-and-jit-compilers |
Baseline benchmark output |
observations/p2-java-assembly-and-task-pinning |
C2-compiled assembly, perf data and taskset runs |
observations/p3-cpuisolation-irqs-memorymgmt |
Runs with CPU isolation, interrupt counts and perf stats |
observations/p7-single-d-array |
Profiles and C2 assembly for the 4D array and the flattened 1D layout |
tools/asm_cfg.py |
Turns a disassembly listing into a Graphviz control-flow graph |
tools/cache_simulator.py |
Set-associative cache simulator that reports hits and misses |
tools/monitor_freq_msr.c |
Reads the APERF/MPERF MSRs to print per-CPU frequency |
tools/monitor_interrupts.py |
Plots local timer interrupts per CPU from /proc/interrupts |
mvn clean package
java -jar target/benchmarks.jar