Utility program for finding and analyzing runtime phases. It uses libscarphase to detect phases, and phase-guided hardware performance counter multiplexing to record the application's phase behavior and performance over time.

See uart/online-phase-detection for more information.

• Linux with perf event support.
• Intel PEBS for sampling branches (works without PEBS, but lower accuracy)

• cmake — CMake build system
• libscarphase — ScarPhase library, pulled in with submodule.
• protobuf — Google data storage library
• boost — portable C++ libraries

• numpy — numeric library
• scipy —
• scikit-learn — machine learning
• matplotlib — graph plotting library
• prettytable — print ascii tables
• progressbar — print progress bar

git clone https://github.com/uart/scarphase.git

cd scarphase
git submodule update --init

cmake .
make

./scarphase -h
usage: ./scarphase <command> [<args>]

Commands:
   profile      Profile stuff
   plot         Plot stuff
   dump         Dump stuff
   show         Show stuff
   simpoint     Find simpoints
   refine       Refine data

See './scarphase help <command>' for more information.

This profiles and finds runtime phases in gcc from SPEC2006, with input 166.i. The profile is saved in gcc.profile and is used in the succeeding examples to plot and dump gcc's phases.

./scarphase profile \
    --scarphase-conf configs/scarphase/example0.conf \
    --counter-conf configs/counters/list0.json \
    --counter-limit=3 \
    gcc.profile \
    -- /path/to/gcc 166.i -o 166.s

• ./scarphase profile - scarphase command
• --scarphase-conf configs/scarphase/example0.conf - contains the configuration settings for the ScarPhase library.
• --counter-conf configs/counters/list0.json - a list of performance counter to sample
• --counter-limit - number of available hardware performance counters
• gcc.profile - output file
• -- gcc 166.i -o 166.s - command to run (use absolute paths)

This plots the performance data and the detected phases from the example above.

./scarphase plot windows -t 0 -c "2" gcc.profile

• ./scarphase plot - scarphase command
• windows - subcommand: plots windows over time
• -t 0 - which thread
• -c "2" - which performance counter to plot
• gcc.profile - profile from example 1

./scarphase dump windows -t 0 gcc.profile
+-----+-----+-----------+--------------+----------------+------------------+--------------+------------------+-------------------+---------------------+---------------+-------------------+-------------------+
| WID | PID |    cycles | instructions |            cpi | cache_references | cache_misses | cache_miss_ratio |   cache_miss_rate | branch_instructions | branch_misses | branch_miss_ratio |  branch_miss_rate |
+-----+-----+-----------+--------------+----------------+------------------+--------------+------------------+-------------------+---------------------+---------------+-------------------+-------------------+
|   0 |   1 |  82651513 |    100003538 | 0.826485888929 |           573485 |        41360 |  0.0721204565071 |  0.00041358536735 |            20520996 |        341234 |   0.0166285301162 |  0.00341221927568 |
|   1 |   1 |  82651513 |    100003538 | 0.826485888929 |           573485 |        41360 |  0.0721204565071 |  0.00041358536735 |            20520996 |        341234 |   0.0166285301162 |  0.00341221927568 |
|   2 |   1 |  82651513 |    100003538 | 0.826485888929 |           573485 |        41360 |  0.0721204565071 |  0.00041358536735 |            20520996 |        341234 |   0.0166285301162 |  0.00341221927568 |
|   3 |   1 |  82651513 |    100003538 | 0.826485888929 |           573485 |        41360 |  0.0721204565071 |  0.00041358536735 |            20520996 |        341234 |   0.0166285301162 |  0.00341221927568 |
|   4 |   2 | 104761301 |    100001455 |  1.04759776745 |           610295 |       525822 |   0.861586609754 |  0.00525814349401 |            20594340 |        217958 |   0.0105833933013 |  0.00217954828757 |
...

• ./scarphase dump - scarphase command
• windows - subcommand: dump windows data over time
• -t 0 - which thread
• gcc.profile - profile from example 1

Output data:

• WID - Window Id (100M instruction per window)
• PID - Phase id
• cycles - executed cycles during the window
• the rest of the performance counter data (configured in list0.json)

• Modeling Performance Variation Due to Cache Sharing Andreas Sandberg, Andreas Sembrant, Erik Hagersten and David Black-Schaffer. In International Symposium on High Performance Computer Architecture (HPCA'13)

• Phase Behavior in Serial and Parallel Applications Andreas Sembrant, David Black-Schaffer and Erik Hagersten. In International Symposium on Workload Characterization (IISWC'12)
• Phase Guided Profiling for Fast Cache Modeling Andreas Sembrant, David Black-Schaffer and Erik Hagersten. In International Symposium on Code Generation and Optimization (CGO'12)
• Low Overhead Instruction-Cache Modeling Using Instruction Reuse Profiles Muneeb Khan, Andreas Sembrant and Erik Hagersten. In International Symposium on Computer Architecture and High Performance Computing (SBAC-PAD'12)
• Power-Sleuth: A Tool for Investigating your Program's Power Behavior Vasileios Spiliopoulos, Andreas Sembrant and Stefanos Kaxiras. In International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS'12)

• Efficient software-based online phase classification Andreas Sembrant, David Eklöv, and Erik Hagersten. In International Symposium on Workload Characterization (IISWC'11)