狠狠撸

Application Pro?ling & Analysis

Rishi Pathak
National PARAM Supercomputing Facility, C-DAC
riship@cdac.in

What is application profiling
? Profiling
– Recording of summary information during execution
? inclusive, exclusive time, # calls, hardware counter statistics, …
– Reflects performance behavior of program entities
? functions, loops, basic blocks
? user-defined “semantic” entities
– Very good for low-cost performance assessment
– Helps to expose performance bottlenecks and hotspots
– Implemented through either
? sampling: periodic OS interrupts or hardware counter traps
? measurement: direct insertion of measurement code

Sampling vs. Instrumentation
Sampling Instrumentation
Overhead Typically about 1% High, may be 500% !

System-wide Yes, profiles all app, drivers, OS functions Just application and
profiling instrumented DLLs

Detect unexpected Yes , can detect other programs using OS No
events resources

Setup None Automatic ins. of data
collection stubs required

Data collected Counters, processor an OS state Call graph , call times,
critical path
Data granularity Assembly level instr., with src line Functions, sometimes
statements
Detects No, Limited to processes , threads Yes – can see algorithm,
algorithmic issues call path is expensive

Profiling Tools 3

Inclusive v/s Exclusive Profiling
int main( )
{ /* takes 100 secs */
f1(); /* takes 20 secs */
/* other work */

/* other work */
}

/* similar for other metrics, such
as hardware performance counters,
etc. */

? Inclusive time for main
– 100 secs

? Exclusive time for main
– 100-20-50-20=10 secs

What are Application Traces
? Tracing
– Recording of information about significant points (events) during program
execution
? entering/exiting code region (function, loop, block, …)
? thread/process interactions (e.g., send/receive message)
– Save information in event record
? timestamp
? CPU identifier, thread identifier
? Event type and event-specific information
– Event trace is a time-sequenced stream of event records
– Can be used to reconstruct dynamic program behavior
– Typically requires code instrumentation

Profiling v/s Tracing

? Profiling
– Summary statistics of performance metrics
? Number of times a routine was invoked
? Exclusive, inclusive time/hpm counts spent executing it
? Number of instrumented child routines invoked, etc.
? Structure of invocations (call-trees/call-graphs)
? Memory, message communication sizes

? Tracing
– When and where events took place along a global timeline
? Time-stamped log of events
? Message communication events (sends/receives) are tracked
? Shows when and from/to where messages were sent
? Large volume of performance data generated usually leads to more
perturbation in the program

The Big Picture

Sampling Instrumentation

Profiling

Analysis Optimization

Measurements - Instrumentation
Instrumentation - Adding measurement probes
to the code to observe its execution
– Can be done on several levels
– Different techniques for different levels
– Different overheads and levels of accuracy with each technique
– No instrumentation: run in a simulator. E.g., Valgrind

? Source code instrumentation
– User added time measurement, etc. (e.g.,
printf(), gettimeofday())
– Many tools expose mechanisms for source code
instrumentation in addition to automatic
instrumentation facilities they offer

– Instrument program phases:
? initialization/main iteration loop/data post processing

? Preprocessor Instrumentation
– Example: Instrumenting OpenMP constructs with
Opari
Orignial Pre- Modified (instrumented)
source code processor source code

– Preprocessor operation
This is used for OpenMP
analysis in tools such as
KoJak/Scalasca/ompP
/* ORIGINAL CODE in parallel region */

Instrumentation
added by Opari

– Example: Instrumenta

? Compiler Instrumentation
– Many compilers can instrument functions
automatically
– GNU compiler flag: -finstrument-
functions
– Automatically calls functions on function
entry/exit that a tool can capture
– Not standardized across compilers, often
undocumented flags, sometimes not available at
all

– GNU compiler example:
void __cyg_profile_func_enter(void *this, void *callsite)
{
/* called on function entry */
}

void __cyg_profile_func_exit(void *this, void *callsite)
{
/* called just before returning from function */
}

? Library Instrumentation:

MPI library interposition

– All functions are available under two names: MPI_xxx and PMPI_xxx, MPI_xxx symbols
are weak, can be over-written by interposition library
– Measurement code in the interposition library measures begin, end, transmitted data,
etc… and calls corresponding PMPI routine.
– Not all MPI functions need to be instrumented

? Binary Instrumentation
– Static binary instrumentation
? LD_PRELOAD(Linux)
? DLL injection(MS)

– Debuggers
– Breakpoints (Software & Hardware)

? Dynamic binary instrumentation

? Injection of instrumentation code into a running process.
? Tools: PIN(Intel), Valgrind

Measurements - Sampling
Using event triggers
– Reoccurring - program counter is sampled many
times
– Histogram of program contexts(CCT)
– Sufficiently large number of samples required
– Uniformity in event triggers wrt execution time

Measurements - Sampling
Event trigger types
? Synchronous
– Initiated by direct program action
– E.g. memory allocation, I/O, and inter-process
communication(including MPI communication)
? Asynchronous
– Not initiated by direct program action
– OS timer interrupt or
– Hardware performance counter events
– E.g. CPU, floating point instructions, clock cycles etc.

Profiling Tools

? Gprof
? Intel VTune
? Scalasca
? HPC Tool Kit
Profiling Tools 18

HPC Tool Kit
? Name: HPCToolkit
? Developer: Rice University
? Website:
– http://hpctoolkit.org

19

HPCToolkit Overview
? Consists of
– hpcviewer
? Sorts by any collected metric, from any processes displayed
? Displays samples at various levels in call hierarchy through “flattening”
? Allows user to focus in on interesting sections of the program through “zooming”
– Hpcprof and hpcprof-mpi
? Correlating dynamic profiles with static source code structure
– hpcrun
? Application profiling using statistical sampling
? Hpcrun-flat – for collection of `flat' profile
– Hpcstruct
? Recovers static program structure such as procedures and loop nests
– Hpctraceviewer
– Hpclink
? For statically-linked executables (e.g. for Cray XT or BG/P)

20

Available Metrics in HPCToolkit
? Metrics, obtained by sampling/profiling
– PAPI Hardware counters
– OS program counters
? Wallclock time (WALLCLK)
– However, can’t get PAPI metrics and Wallclock time in a single run
? Derived metrics
– Combination of existing metrics created by specifying a mathematical formula in an XML
configuration file.
? Source Code Correlation
– Metrics reflect exclusive time spent in function based on counter overflow events
– Metrics correlated at the source line level and the loop level
– Metrics are related back to source code loops (even if code has been significantly altered
by optimization) (“bloop”)

21

hpcviewer Views

? Calling context view
– top-down view shows dynamic calling contexts in which costs were
incurred

? Caller’s view
– bottom-up view apportions costs incurred in a routine to the
routine’s dynamic calling contexts

? Flat view
– aggregates all costs incurred by a routine in any context and shows
the details of where they were incurred within the routine

hpctraceviewer Views

? Trace view (left, top)
– Time on the horizontal axis
– Process (or thread) rank on the vertical axis
? Depth view (left, bottom) & Summary view
– Call-path/time view for the process rank selected
? Call view (right, top)
– Current call path depth that defines the hierarchical slice
shown in the Trace View
– Actual call path for the point selected by the Trace View's
crosshair

Hpctraceviewer user interface - DEMO

Call Path Profiling: Costs in Context

Event-based sampling method for performance measurement
? When a profile event occurs, e.g. a timer expires
– determine context in which cost is incurred
? unwind call stack to determine set of active procedure frames
– attribute cost of sample to PC in calling context
? Benefits
– monitor unmodified fully optimized code
– language independent – C/C++, Fortran, assembly code, …
– accurate
– low overhead (1K samples per second has ~ 3-5% overhead)

Demo for :
? Hpcrun (list events & proifiling)
? Hpcstruct
? hpcprof

PAPI
? Performance Application Programming
Interface
– The purpose of the PAPI project is to design, standardize
and implement a portable and efficient API to access the
hardware performance monitor counters found on most
modern microprocessors.
? Parallel Tools Consortium project started in 1998
? Developed by University of Tennessee, Knoxville
? http://icl.cs.utk.edu/papi/

PAPI - Support
? Unix/Linux
– Perfctr kernel patch for kernel < 2.6.30
– Perf package for kernel >= 2.6.30

PAPI - Implementation
3rd Party and GUI Tools

Portable PAPI High Level
Layer PAPI Low Level

PAPI Machine Dependent Substrate
Machine Kernel Extension
Specific
Layer Operating System

Hardware Performance Counters

PAPI - Hardware Events
? Preset Events(Platform neutral)
– Standard set of over 100 events for application performance tuning
– No standardization of the exact definition
– Mapped to either single or linear combinations of native events on each
platform
– Use papi_avail utility to see what preset events are available on a given
platform
– PAPI_TOT_INS
? Native Events(Platform dependent)
– Any event countable by the CPU
– Same interface as for preset events
– Use papi_native_avail utility to see all available native events
– L3_MISSES
? Use papi_event_chooser utility to select a compatible set of events

PAPI events demo
papi_avail
papi_native_avail
papi_event_chooser
Availability for QPI h/w performance
counters using /sbin/lspci

PAPI-Derived Metrics
Metric Formula
Instructions
Graduated instructions per cycle PAPI_TOT_INS/PAPI_TOT_CYC
Issued instructions per cycle PAPI_TOT_IIS/PAPI_TOT_CYC

Graduated floating point instructions per cycle PAPI_FP_INS/PAPI_TOT_CYC

Percentage floating point instructions PAPI_FP_INS/PAPI_TOT_INS

Ratio of graduated instructions to issued instructions PAPI_TOT_INS/PAPI_TOT_IIS

Percentage of cycles with no instruction issue 100.0 * (PAPI_STL_ICY/PAPI_TOT_CYC)

Data references per instruction PAPI_L1_DCA/PAPI_TOT_INS

Ratio of floating point instructions to L1 data cache accesses PAPI_FP_INS/PAPI_L1_DCA

Ratio of floating point instructions to L2 cache accesses (data) PAPI_FP_INS/PAPI_L2_DCA

Issued instructions per L1 instruction cache miss PAPI_TOT_IIS/PAPI_L1_ICM

Graduated instructions per L1 instruction cache miss PAPI_TOT_INS/PAPI_L1_ICM

L1 instruction cache miss ratio PAPI_L2_ICR/PAPI_L1_ICR

Metric Formula
Cache & Memory Hierarchy
Graduated loads & stores per cycle PAPI_LST_INS/PAPI_TOT_CYC

Graduated loads & stores per floating point instruction PAPI_LST_INS/PAPI_FP_INS

L1 cache line reuse (data) ((PAPI_LST_INS - PAPI_L1_DCM) / PAPI_L1_DCM)

L1 cache data hit rate 1.0 - (PAPI_L1_DCM/PAPI_LST_INS)
L1 data cache read miss ratio PAPI_L1_DCM/PAPI_L1_DCA

L2 cache line reuse (data) ((PAPI_L1_DCM - PAPI_L2_DCM) / PAPI_L2_DCM)

L2 cache data hit rate 1.0 - (PAPI_L2_DCM/PAPI_L1_DCM)
L2 cache miss ratio PAPI_L2_TCM/PAPI_L2_TCA

L3 cache line reuse (data) ((PAPI_L2_DCM - PAPI_L3_DCM) / PAPI_L3_DCM)

L3 cache data hit rate 1.0 - (PAPI_L3_DCM/PAPI_L2_DCM)
L3 data cache miss ratio PAPI_L3_DCM/PAPI_L3_DCA
L3 cache data read ratio PAPI_L3_DCR/PAPI_L3_DCA
L3 cache instruction miss ratio PAPI_L3_ICM/PAPI_L3_ICR
((PAPI_Lx_TCM * Lx_linesize) / PAPI_TOT_CYC)
Bandwidth used (Lx cache)
* Clock(MHz)

Metric Formula
Branching

Ratio of mispredicted to correctly predicted branches PAPI_BR_MSP/PAPI_BR_PRC

Processor Stalls

Percentage of cycles waiting for memory access 100.0 * (PAPI_MEM_SCY/PAPI_TOT_CYC)

Percentage of cycles stalled on any resource 100.0 * (PAPI_RES_STL/PAPI_TOT_CYC)

Aggregate Performance

MFLOPS (CPU cycles) (PAPI_FP_INS/PAPI_TOT_CYC) * Clock(MHz)

MFLOPS (effective) PAPI_FP_INS/Wallclock time

MIPS (CPU cycles) (PAPI_TOT_INS/PAPI_TOT_CYC) * Clock(MHz)

MIPS (effective) PAPI_TOT_INS/Wallclock time

Processor utilization (PAPI_TOT_CYC*Clock) / Wallclock time

http://perfsuite.ncsa.illinois.edu/psprocess/metrics.shtml

Component PAPI (PAPI-C)
? Goals:
– Support for simultaneous access to on- and off-processor
counters
– Isolation of hardware dependent code in a separable
‘substrate’ module
– Extension of platform independent code to support
multiple simultaneous substrates
– API calls to support access to any of several substrates

? Released in PAPI 4.0

Extension to PAPI to
Support Multiple Substrates

PAPI High Level

PAPI Low Level
Portable
Layer Hardware Independent Layer

PAPI Machine Dependent Substrate PAPI Machine Dependent Substrate
Machine Kernel Extension Kernel Extension
Specific
Operating System Operating System
Layer
Hardware Performance Counters Off-Processor Hardware Counters

High-level tools that use PAPI
? TAU (U Oregon)

? HPCToolkit (Rice Univ)

? KOJAK (UTK, FZ Juelich)

? PerfSuite (NCSA)

? SCALASCA

? Open|Speedshop (SGI)

? Intel Vtune

? Hpcviewer demo (trace, flops and clock cycles)
– Nek5000(CFD solver using spectral element
method)
– Xhpl(Linpack)

狠狠撸

HPC Application Profiling and Analysis

More Related Content

HPC Application Profiling and Analysis