Profiling

pprof

pprof helps diagnose which functions use most of the CPU resources. Note that this works better by compiling lib with a test file rather than linking against it.

Run

• compile with -g and link with -lprofiler
• run with LDPRELOAD=/usr/lib/libprofiler.so CPUPROFILE=cpu.profile ./bin
• run google-pprof --text cpu.profile ./bin (switch to --gv, --pdf, --svg, ... depending on what output you prefer)

cachegrind

cachegrind helps diagnose data & instructions cache misses. Note that this works better by compiling lib with a test file rather than linking against it.

Run

• compile with -g
• run valgrind --tool=cachegrind ./bin
• run cg_annotate cachegrind.out.xxxx to get a global report by file
• run cg_annotate cachegrind.out.xxxx /abs/path/to/file.cpp to get a line-by-line report for a specific file
• run cg_annotate cachegrind.out.xxxx --auto=yes to get a full line-by-line report for all files

Output

Interpretation

• I: Instruction Cache
• I1: L1 Instruction Cache
• LLi: Last Level Instruction Cache (L2 on valgrind, simulates L3/L4/... basically last level for machines with more than 2 cache levels).
• D: Data cache
• D1: L1 Data cache
• LLd: Last Level Data Cache (same as LLi)
• r: read
• w: write
• mr: miss on read
• mw: miss on write
• Dmr, D1mr, LLdmr, ... combination of cache type and action type (e.g: D1mr is miss on read for L1 Data cache)

Misses are easy to spot, especially on the line-by-line detail.

If one line has a high rate of data misses, it may be interesting to reorganize the way the data is accessed to take profit of cache locality.

Not sure how to address instructions misses.

callgrind

callgrind helps diagnose which instructions are executed the most. Note that this works better by compiling lib with a test file rather than linking against it.

Run

• compile with -g
• run valgrind --tool=callgrind ./bin
• run callgrind_annotate callgrind.out.xxxx to get a global report by file
• run callgrind_annotate callgrind.out.xxxx --auto=yes to get a full line-by-line report for all files

kcachegrind

kcachegrind provides is an alternative to cg_annotate and callgrind_annotate to navigate through the reports. It provides a graphic interface that can be much more convenient depending on the context.

Run

• run kcachegrind callgrind.out.xxx or kcachegrind cachegrind.out.xxx

Approach

Top-down approach can be interesting to focus on the relevant areas.

Typically:

• identify slow extended APIs based on google tests runtime.
• identify slow components with pprof
• once identified, find root cause using cachegrind and callgrind