Screencast link: https://www.youtube.com/watch?v=G32QMsgr-jM

Last Class

• Measuring wall-time
• Calculating FLOP/s and MOP/s
• Compiler optimization flags: -Ox, -march=native, -g
• Brief overview of different kinds of parallelism:
  • instruction level (vectorization, pipelining)
  • shared memory (OpenMP)
  • distributd memory (MPI)
• Tools: ssh, modules

Definitions

• Arithmetic intensity
• Latency
• Bandwidth
• Cache
• Cache line
• Virtual memory, Translation Lookaside Buffer (TLB), Memory pages, Page fault, First touch

Valgrind

--tool=memcheck (default)

• Out-of-bound array access
• Uninitialized variables (--track-origins=yes)
• Memory leaks (--leak-check=full)
• Attach GDB (--vgdb-error=0)

--tool=callgrind (for profiling/timing)

• kcachegrind
• Optional agruments --dump-instr=yes --collect-jumps=yes
• cg_annotate --auto=yes
• On Mac: brew install qcachegrind --with-graphviz

--tool=cachegrind (for cache misses)

• From callgrind --tool=callgrind --cache-sim=yes

References

• http://igoro.com/archive/gallery-of-processor-cache-effects/
• https://people.freebsd.org/~lstewart/articles/cpumemory.pdf

Takeaway

• Memory allocations are expensive (i.e. cost of first touch not malloc)
• Keep most frequently used data in registers, then L1 cache, L2 cache, L3 cache and main memory
• Regular memory access patterns (implicit prefetch) or explicit prefetch to hide memory latency
• Valgrind to debug memory errors, profile code and find cache misses