The initial task in a performance analysis is to find out where the runtime is spent. You want to focus on the regions of the code that consume significant runtime. Runtime profilers allow to measure the runtime distribution across the application code. Two profiler variants exist: Instrumentation based and sampling based. Instrumentation based profilers insert function calls to measure the time at points in the program. Additional tasks performed are e.g. determining the function call stack. While it is possible to insert instrumentation calls on the binary level usually the instrumentation calls are added during compilation. The standard tool is gprof and almost any compiler supports to instrument the code for gprof. Statistical sampling based profiling on the other hand is probing the program call stack at regular intervals. A widespread tool for sampling based profiling is the perf tool which builds on the builtin profiling infrastructure in recent Linux kernels. Both approaches have advantages and disadvantages: Instrumentation produces more accurate results but introduces more overhead and sampling has less overhead but produce less accurate results. Special care is necessary for runtime profiling of parallel (OpenMP or MPI) applications and C++ programs (due to name mangling resolving issues).
How to use gprof
The first step is to compile and link the program with profiling enabled. For most compilers this is done by setting the -pg flag. For Intel compilers it is important to set the optimisation flag afterward as the default optimisation level is set to -O0 when profiling is enabled.
Example for build options with Intel tool chain:
icc -pg -O3 -c myfile1.c icc -pg -O3 -c myfile2.c icc -o a.out -pg myfile1.o myfile2.o
When generated application is executed it generates a gmon.out file containing the profiling output. To analyse the profile the tool gprof is used with the executable as argument:
Refer to the gprof man page for additional command line options. The result will be printed on stdout. It is recommended to either redirect the output to a file or use a pipe to the less pager command.
gprof ./a.out | less
The default output consists of three parts: A flat profile, the call graph, and an alphabetical index of routines. For most purposes the flat profile is what you are looking for.
Example output of the flat profile for the Mantevo miniMD proxy app:
Each sample counts as 0.01 seconds. % cumulative self self total time seconds seconds calls s/call s/call name 66.86 26.14 26.14 502 0.05 0.05 ForceLJ::compute(Atom&, Neighbor&, Comm&, int) 30.77 38.17 12.03 26 0.46 0.46 Neighbor::build(Atom&) 1.43 38.73 0.56 1 0.56 38.46 Integrate::run(Atom&, Force*, Neighbor&, Comm&, Thermo&, Timer&) 0.36 38.87 0.14 2850 0.00 0.00 Atom::pack_comm(int, int*, double*, int*) 0.15 38.93 0.06 2850 0.00 0.00 Atom::unpack_comm(int, int, double*) 0.13 38.98 0.05 26 0.00 0.00 Atom::pbc() 0.10 39.02 0.04 __intel_ssse3_rep_memcpy 0.08 39.05 0.03 25 0.00 0.00 Atom::sort(Neighbor&) 0.08 39.08 0.03 1 0.03 0.03 create_atoms(Atom&, int, int, int, double) 0.05 39.10 0.02 26 0.00 0.00 Comm::borders(Atom&) 0.00 39.10 0.00 1221559 0.00 0.00 Atom::pack_border(int, double*, int*) 0.00 39.10 0.00 1221559 0.00 0.00 Atom::unpack_border(int, double*) 0.00 39.10 0.00 131072 0.00 0.00 Atom::addatom(double, double, double, double, double, double) 0.00 39.10 0.00 1025 0.00 0.00 Timer::stamp(int) 0.00 39.10 0.00 502 0.00 0.00 Thermo::compute(int, Atom&, Neighbor&, Force*, Timer&, Comm&) 0.00 39.10 0.00 500 0.00 0.00 Timer::stamp() 0.00 39.10 0.00 475 0.00 0.00 Comm::communicate(Atom&) 0.00 39.10 0.00 26 0.00 0.00 Comm::exchange(Atom&) 0.00 39.10 0.00 25 0.00 0.00 Timer::stamp_extra_stop(int) 0.00 39.10 0.00 25 0.00 0.00 Timer::stamp_extra_start() 0.00 39.10 0.00 25 0.00 0.00 Neighbor::binatoms(Atom&, int) 0.00 39.10 0.00 7 0.00 0.00 Timer::barrier_stop(int) 0.00 39.10 0.00 1 0.00 0.00 create_box(Atom&, int, int, int, double) 0.00 39.10 0.00 1 0.00 0.00 create_velocity(double, Atom&, Thermo&)
The output is sorted according to the total time spent in it. The interesting columns are self seconds (the time spent in the routine itself), calls (how often it was called) and self s/call (how much time was spent per call).
How to use perf
Runtime profiling with perf is very simple. You execute your executable wrapped with the perf call:
perf record ./miniMD
After the application finished the results can be analysed with
which opens a ncurses based presentation of the results:
Samples: 30K of event 'cycles:uppp', Event count (approx.): 20629160088 Overhead Command Shared Object Symbol 64.19% miniMD-ICC miniMD-ICC [.] ForceLJ::compute 31.54% miniMD-ICC miniMD-ICC [.] Neighbor::build 1.47% miniMD-ICC miniMD-ICC [.] Integrate::run 0.67% miniMD-ICC [kernel] [k] irq_return 0.40% miniMD-ICC miniMD-ICC [.] Atom::pack_comm 0.35% mpiexec [kernel] [k] sysret_check 0.21% miniMD-ICC miniMD-ICC [.] create_atoms 0.18% miniMD-ICC miniMD-ICC [.] Atom::unpack_comm 0.15% miniMD-ICC [kernel] [k] sysret_check 0.15% miniMD-ICC miniMD-ICC [.] Comm::borders 0.10% miniMD-ICC miniMD-ICC [.] __intel_ssse3_rep_memcpy 0.09% miniMD-ICC miniMD-ICC [.] Atom::sort 0.07% miniMD-ICC miniMD-ICC [.] Neighbor::binatoms 0.05% mpiexec [kernel] [k] irq_return 0.04% miniMD-ICC miniMD-ICC [.] Atom::pbc 0.03% miniMD-ICC miniMD-ICC [.] Atom::unpack_border 0.03% miniMD-ICC miniMD-ICC [.] Atom::addatom 0.02% miniMD-ICC miniMD-ICC [.] Atom::pack_border 0.02% hydra_pmi_proxy [kernel] [k] sysret_check 0.01% miniMD-ICC miniMD-ICC [.] create_velocity 0.01% mpiexec libc-2.17.so [.] vfprintf 0.01% miniMD-ICC ld-2.17.so [.] _dl_lookup_symbol_x 0.01% miniMD-ICC ld-2.17.so [.] do_lookup_x 0.01% hydra_bstrap_pr [kernel] [k] irq_return 0.01% hydra_pmi_proxy [kernel] [k] irq_return 0.01% hydra_bstrap_pr [kernel] [k] sysret_check 0.01% miniMD-ICC libmpi.so.12.0.0 [.] MPIR_T_CVAR_REGISTER_impl 0.01% miniMD-ICC libc-2.17.so [.] getenv 0.01% miniMD-ICC libmpi.so.12.0.0 [.] MPL_wtime
As can be seen the result is similar but not identical to gprof. Perf also reports on time spent in external entities as the kernel, mpiexec and system libraries.
Runtime profiling for parallel applications
gprof as well as perf are designed to be used for single process applications. While there are possibilities to use gprof in a MPI setting (see the LRZ documentation) a lightweight alternative to complex full trace tools is the command line version of Intel Amplifier. It is also using the builtin perf infrastructure in the Linux kernel.
Example usage with MPI:
mpirun -np 1 amplxe-cl -collect hotspots -result-dir myresults -- ../miniMD-ICC --half_neigh 0
The analysis is printed on stdout:
Elapsed Time: 8.650s CPU Time: 8.190s Effective Time: 8.190s Idle: 0.020s Poor: 8.170s Ok: 0s Ideal: 0s Over: 0s Spin Time: 0s Overhead Time: 0s Total Thread Count: 2 Paused Time: 0s Top Hotspots Function Module CPU Time --------------------------- ---------- -------- ForceLJ::compute_fullneigh miniMD-ICC 4.940s Neighbor::build miniMD-ICC 2.820s Integrate::finalIntegrate miniMD-ICC 0.100s Integrate::initialIntegrate miniMD-ICC 0.060s __intel_ssse3_rep_memcpy miniMD-ICC 0.040s [Others] N/A 0.230s Effective Physical Core Utilization: 4.9% (0.976 out of 20) | The metric value is low, which may signal a poor physical CPU cores | utilization caused by: | - load imbalance | - threading runtime overhead | - contended synchronization | - thread/process underutilization | - incorrect affinity that utilizes logical cores instead of physical | cores | Explore sub-metrics to estimate the efficiency of MPI and OpenMP parallelism | or run the Locks and Waits analysis to identify parallel bottlenecks for | other parallel runtimes. | Effective Logical Core Utilization: 2.4% (0.976 out of 40) | The metric value is low, which may signal a poor logical CPU cores | utilization. Consider improving physical core utilization as the first | step and then look at opportunities to utilize logical cores, which in | some cases can improve processor throughput and overall performance of | multi-threaded applications. | Collection and Platform Info Application Command Line: ../miniMD-ICC "--half_neigh" "0" Operating System: 3.10.0-957.10.1.el7.x86_64 NAME="CentOS Linux" VERSION="7 (Core)" ID="centos" ID_LIKE="rhel fedora" VERSION_ID="7" PRETTY_NAME="CentOS Linux 7 (Core)" ANSI_COLOR="0;31" CPE_NAME="cpe:/o:centos:centos:7" HOME_URL="https://www.centos.org/" BUG_REPORT_URL="https://bugs.centos.org/" CENTOS_MANTISBT_PROJECT="CentOS-7" CENTOS_MANTISBT_PROJECT_VERSION="7" REDHAT_SUPPORT_PRODUCT="centos" REDHAT_SUPPORT_PRODUCT_VERSION="7" Computer Name: e1125 Result Size: 3 MB Collection start time: 13:11:20 02/04/2019 UTC Collection stop time: 13:11:29 02/04/2019 UTC Collector Type: Driverless Perf per-process counting,User-mode sampling and tracing CPU Name: Intel(R) Xeon(R) E5/E7 v2 Processor code named Ivytown Frequency: 2.200 GHz Logical CPU Count: 40
This was also the only tool that resolved the inlined C++ routine. There are many more collect modules available including hardware performance monitoring metrics. Please refer to the Intel documentation for more detailed information.