IBM Hardware Performance Monitor (hpm)

1
IBM Hardware Performance Monitor (hpm)

• NPACI Parallel Computing Institute August, 2002

2
What is Performance?

• - Where is time spent and how is time spent?
• MIPS Millions of Instructions Per Second
• MFLOPS Millions of Floating-Point Operations
  Per Second
• Run time/CPU time

3
What is a Performance Monitor?

• Provides detailed processor/system data
• Processor Monitors
• Typically a group of registers
• Special purpose registers keep track of
  programmable events
• Non-intrusive counts result in accurate
  measurement of processor events
• Typical Events counted are Instruction, floating
  point instruction, cache misses, etc.
• System Level Monitors
• Can be hardware or software
• Intended to measure system activity
• Examples
• bus monitor measures memory traffic, can analyze
  cache coherency issues in multiprocessor system
• Network monitor measures network traffic, can
  analyze web traffic internally and externally

4
Hardware Counter Motivations

• To understand execution behavior of application
  code
• Why not use software?
• Strength simple, GUI interface
• Weakness large overhead, intrusive, higher
  level, abstraction and simplicity
• How about using a simulator?
• Strength control, low-level, accurate
• Weakness limit on size of code, difficult to
  implement, time-consuming to run
• When should we directly use hardware counters?
• Software and simulators not available or not
  enough
• Strength non-intrusive, instruction level
  analysis, moderate control, very accurate, low
  overhead
• Weakness not typically reusable, OS kernel
  support

5
Ptools Project

• PMAPI Project
• Common standard API for industry
• Supported by IBM, SUN, SGI, COMPAQ etc
• PAPI Project
• Standard application programming interface
• Portable, available through a module
• Can access hardware counter info
• HPM Toolkit
• Easy to use
• Doesnt effect code performance
• Use hardware counters
• Designed specifically for IBM SPs and Power

6
Problem Set

• Should we collect all events all the time?
• Not necessary and wasteful
• What counts should be used?
• Gather only what you need

• Cycles
• Committed Instructions
• Loads
• Stores
• L1/L2 misses
• L1/L2 stores

• Committed fl pt instr
• Branches
• Branch misses
• TLB misses
• Cache misses

7
POWER3 Architecture

8
IBM HPM Toolkit

• High Performance Monitor
• Developed for performance measurement of
  applications running on IBM Power3 systems. It
  consists of
• An utility (hpmcount)
• An instrumentation library (libhpm)
• A graphical user interface (hpmviz).
• Requires PMAPI kernel extensions to be loaded
• Works on IBM 630 and 604e processors
• Based on IBMs PMAPI low level interface

9
HPM Count

• Utilities for performance measurement of
  application
• Extra logic inserted to the processor to count
  specific events
• Updated at every cycle
• Provide a summary output at the end of the
  execution
• Wall clock time
• Resource usage statistics
• Hardware performance counters information
• Derived hardware metrics
• Serial/Parallel, Gives each performance numbers
  for each task

10
HPM Usage HW Event Categories

• EVENT SET 1
• Cycles
• Inst. Completed
• TLB misses
• Stores completed
• Loads completed
• FPU0 ops
• FPU1 ops
• FMAs executed 

EVENT SET 2 Cycles Inst. Completed TLB
misses Stores dispatched L1 store misses Loads
dispatched L1 load misses LSU idle 
EVENT SET 3 Cycles Inst. dispatched Inst.
Completed Cycles w/  0 inst. completed I cache
misses FXU0 ops FXU1 ops FXU2 ops
EVENT SET 4 Cycles Loads dispatched L1 load
misses L2 load misses Stores dispatched L2 store
misses Comp. unit waiting on load LSU idle
floating point performance and usage of floating
point units
performance and usage of level 1 instruction cache
usage of level 2 data cache and branch prediction
data locality and usage of level 1 data cache
11
HPM for Whole Program using HPMCOUNT

• Installed in /usr/local/apps/hpm,
  /usr/local/apps/HPM_V2.3
• Environment setting
• setenv LIBHPM_EVENT_SET 1 (2,3,4)
• setenv MP_LABELIO YES -gt to correlate each
  line of output with corresponding task
• setenv MP_STDOUTMODE -gttaskID(e.g. 0) to
  discard output from other tasks
• Usage
• poe hpmcount ./a.out -nodes 1 -tasks_per_node
  1 -rmpool 1 -s ltsetgt -e ev,ev -h
• -h displays a help message
• -e ev0,ev1, list of event numbers, separated by
  commas
• evltigt corresponds to event selected for counter
  ltIgt
• -s predefined set of envets

12
Derived Hardware Metrics

• Hardware counters provide only raw counts
• 8 counters on Power3
• Enough info for generation of derived metrics on
  each execution
• Derived Metrics
• Floating point rate
• Computational Intensity
• Instruction per load / store
• Load/store per data cache misses
• Cache hit rate
• Loads per load miss
• Stores per store miss
• Loads per TLB miss
• FMA
• Branches Mispredicted

13
HPMCOUNT Output (Event1) Resource Usage
Statistics

• Total execution time of instrumented code (wall
  time) 6.218496 seconds
• Total amount of time in user mode
  5.860000 seconds
• Total amount of time in system mode
  3.120000 seconds
• Maximum resident set size
  23408 Kbytes
• Average shared memory use in text segment
  97372 Kbytessec
• Average unshared memory use in data segment
  13396800 Kbytessec
• Number of page faults without I/O activity
  5924
• Number of page faults with I/O activity
  12
• Number of times process was swapped out 0
• Number of times file system performed INPUT
  0
• Number of times file system performed OUTPUT 0
• Number of IPC messages sent
  0
• Number of IPC messages received
  0
• Number of signals delivered
  0
• Number of voluntary context switches
  2840
• Number of involuntary context switches
  27740

14
HPMCOUNT (Event1 continued) Resource statistics

• Instrumented section 1 - Label ALL - process
  1
• file swim_omp.f, lines 89 lt--gt 189
• Count 1
• Wall Clock Time
  6.216718 seconds
• Total time in user mode
  5.35645462067771 seconds
• Exclusive duration 0.012166 seconds
• PM_CYC (Cycles)
  2008608171
• PM_INST_CMPL (Instructions completed)
  1891769436
• PM_TLB_MISS (TLB misses)
  2374441
• PM_ST_CMPL (Stores completed)
  274169278
• PM_LD_CMPL (Loads completed)
  672275023
• PM_FPU0_CMPL (FPU 0 instructions)
  528010431
• PM_FPU1_CMPL (FPU 1 instructions)
  245779486
• PM_EXEC_FMA (FMAs executed)
  270299532

15
Timers

• Time usually reports three metrics
• User Time
• The time used by your code on CPU, also CPU time
• Total time in user mode Cycles/Processor
  Frequency
• System Time
• The time used by your code running kernel code
  (doing I/O, writing to disk, or printing to the
  screen etc).
• It is worth to minimize the system time, by
  speeding up the disk I/O, doing I/O in parallel,
  or doing I/O in background while your CPU
  computes in the foreground
• Wall Clock time
• Total execution time, the combination of the time
  1 and 2 plus the time spent idle (waiting for
  resources)
• In parallel performance tuning, only wall clock
  time counts
• Interprocessor communication consumes a
  significant amount of your execution time
  (user/system time usually dont account for it),
  need to rely on wall clock time for all the time
  consumed by the job

16
Floating Point Measures

• PM_FPU0_CMPL (FPU 0 instructions)
• The POWER3 processor has two Floating Point Units
  (FPU) which operate in parallel. Each FPU can
  start a new instruction at every cycle. This
  counter shows the number of floating point
  instructions that have been executed by the first
  FPU.
• PM_FPU1_CMPL (FPU 1 instructions)
• This counter shows the number of floating point
  instructions (add, multiply, subtract, divide,
  multiply add) that have been processed by the
  second FPU.
• PM_EXEC_FMA (FMAs executed)
• This is the number of Floating point Multiply
  Add (FMA) instructions. This instruction does a
  computation of following type x s a b So
  two floating point operations are done within one
  instruction. The compiler generate this
  instruction as often as possible to speed up the
  program. But sometimes additional manual
  optimization is necessary to replace single
  multiply instructions and corresponding add
  instructions by one FMA.

17
HPMCOUNT (Event1 continued)

• Utilization rate
  86.162
• TLB misses per cycle
  0.118
• Estimated latency from TLB misses
  4.432 sec
• Avg number of loads per TLB miss
  283.130
• Load and store operations
  946.444 M
• Instructions per load/store
  1.999
• MIPS
  304.304
• Instructions per cycle
  0.942
• HW Float points instructions per Cycle
  0.385
• Floating point instructions FMAs
  1044.089 M
• Float point instructions FMA rate
  167.949 Mflip/s
• FMA percentage
  51.777
• Computation intensity
  1.103

18
Total Flop Rate

• Float point instructions FMA rate
• This is the most often mentioned performance
  index, the MFlops rate.
• The peak performance of the POWER3-II processor
  is 1500 MFlops. (375 MHZ clock x 2 FPUs x 2
  Flops/FMA instruction).
• Many applications do not reach more than 10
  percent of this peak performance.
• Average number of loads per TLB miss
• This value is the ratio PM_LD_CMPL / PM_TLB_MISS.
  Each time after a TLB miss has been processed,
  fast access to a new page of data is possible.
  Small values for this metric indicate that the
  program has a poor data locality, a redesign of
  the data structures in the program may result in
  significant performance improvements.
• Computation intensity
• Computational intensity is the ratio of Load and
  store operations and Floating point operations

19
HPM for Part of Program using LIBHPM

• Instrumentation of performance library for
  performance measurement of Fortran, C and C
  applications
• Collects information and performs summarization
  during run-time, generate performance file for
  each task
• Use the same set of hardware counters events used
  by hpmcount
• User can specify an event set with the file
  libHPMevents
• For each instrumented point in a program, libhpm
  provides output
• Total count
• Total duration (wall clock time)
• Hardware performance counters information
• Hardware derived metrics
• Supports
• multiple instrumentation points, nested
  instrumentation
• OpenMP and thread applications
• Multiple calls to an instrumented point

20
LIBHPM Functions

• C C
• hpmInit(taskID)
• hpmTerminate(taskID)
• hpmStart(instID)
• hpmStop(instID)
• hpmTstart(instID)
• hpmTstop(instID)

• Fortran
• f_hpminit(taskID)
• f_hpmterminate(taskID)
• f_ hpmstart(instID)
• f_ hpmstop(instID)
• f_ hpmtstart(instID)
• f_ hpmtstop(instID)

21
Using LIBHPM - C

• Declaration
• include libhpm.h
• C usage
• MPI_Comm_rank( MPI_COMM_WORLD, taskID)
• hpmInit(taskID,hpm_test)
• hpmStart(1,outer call)
• code segment to be timed
• hpmStop( 1)
• hpmTerminate(taskID)
• Compilation
• mpcc_r -I/usr/local/apps/HPM_V2.3/include -O3
  -lhpm_r -lpmapi -lm -qarchpwr3 -qstrict
  -qsmpomp -L/usr/local/apps/HPM_V2.3/lib
  hpm_test.c -o hpm_test.x

22
Using LIBHPM - Fortran

• Declaration
• include f_hpm.h
• Fortran usage
• CALL MPI_COMM_RANK( MPI_COMM_WORLD, taskid,
  ierr )
• call f_hpminit(taskID)
• call f_hpmstart(instID)
• code segment to be timed
• call f_hpmstop(instID)
• call f_hpmterminate(taskID)
• CALL MPI_FINALIZE(ierr)
• Compilation
• mpxlf_r -I/usr/local/apps/HPM_V2.3/include
  -qsuffixcppf -O3 -qarchpwr3 -qstrict -qsmpomp
  -L/usr/local/apps/HPM_V2.3/lib -lhpm_r -lpmapi
  -lm hpm_test.f -o hpm_test.x

23
Using LIBHPM - Threads

• call f_hpminit(taskID)
• //do
• call f_hpmtstart(10)
• do_work
• call f_hpmtstop(10)
• end //do
• //do
• call f_hpmtstart(20my_thread_ID )
• do_work
• call f_hpmtstop(20my_thread_ID )
• end //do
• call f_hpmterminate(taskID)

24
HPM Example Code in C
include ltmpi.hgt include ltstdio.hgt include
"libhpm.h" define n 10000 main(int argc, char
argv) int taskID,i,numprocs double
an,bn,cn MPI_Init(argc,argv) MPI_Comm_si
ze(MPI_COMM_WORLD,numprocs) MPI_Comm_rank(MPI_CO
MM_WORLD,taskID) hpmInit(taskID,"hpm_test") hpm
Start(1,section 1") for(i1iltn1i) aii
bin-1 hpmStop(1)
hpmStart(2, "section 2") for(i2iltn1i)
ciaibiai/bi hpmStop(2) hpmTermin
ate(taskID) MPI_Finalize()

25
HPM Example Code in Fortran

• program hpm_test
• parameter (n10000)
• integer taskID,ierr,numtasks
• dimension a(n),b(n),c(n)
• include "mpif.h"
• include "f_hpm.h"
• call MPI_INIT(ierr)
• call MPI_COMM_RANK(MPI_COMM_WORLD,taskID,ier
  r)
• call MPI_COMM_SIZE(MPI_COMM_WORLD,numtasks,i
  err)
• call f_hpminit(taskID,"hpm_test")
• call f_hpmstart(5,section1")
• do i1,n
• a(i)real(i)
• b(i)real(n-i)
• enddo

call f_hpmstop(5) call
f_hpmterminate(taskID) call
MPI_FINALIZE(ierr) end
26
Compiling and Linking

• FF mpxlf_r
• HPM_DIR /usr/local/apps/HPM_V2.3
• HPM_INC -I(HPM_DIR)/include
• HPM_LIB -L(HPM_DIR)/lib -lhpm_r -lpmapi -lm
• FFLAGS -qsuffixcppf -O3 -qarchpwr3 -qstrict
  -qsmpomp
• Note -qsuffixcppf is only required for
  Fortran code with .f
• hpm_test.x hpm_test.f
• (FF) (HPM_INC) (FFLAGS) hpm_test.f (HPM_LIB)
  -o hpm_test.x

27
HPMVIZ

• takes as input the performance files generated by
  libhpm
• Usage
•  gt hpmviz ltperformance files(.viz)gt
• define a range of values considered satisfactory
• Red below predefined as minimum recommended
  value
• Green above the threshold value
• HPMVIZ left pane of the window
• displays for each instrumented point, identified
  by its label, the inclusive duration, exclusive,
  and count.
• HPMVIZ right pane of the window
• shows the corresponding source code which can be
  edited and saved.
• The metrics windows
• display the task ID, Thread ID, count, exclusive
  duration, inclusive duration, and the derived
  hardware metrics.

28
HPMVIZ
29
IBM SP HPM Toolkit Summary

• A complete problem set
• Derived metrics
• Analysis of error message
• Analyze derived metrics
• HPMCOUNT very accurate with low overhead,
  non-intrusive, general view for whole program
• LIBHPM same sets as hpmcount, for part of
  program
• HPMVIZ easier to view the hardware counters
  information and derived metrics

30
HPM References

• HPM README file in /usr/local/apps/HPM_V2.3
• Online Documentation
• http//www.sdsc.edu/SciApps/IBM_tools/hpm.html

31
Lab Session for HPMEnvironment Setup

• Setup for running X-windows applications on PCs
• 1. Login to b80login.sdsc.edu using CRT
  (located in Applications common).
• 2. Launch Exceed (located in either
  Applications (Common) or as a shortcut on your
  desktop called "Humming Bird".
• 3. set your environment, for csh
• setenv DISPLAY t-wolf.sdsc.edu0.0
• where "t-wolf for example is the name
  of the PC you are using
• 4. copy files from /work/Training/HPM_Training
  directory into your own working space.
• create a directory to work with HPM
• mkdir HPM
• change directories into new directory
• cd HPM
• copy files into new directory
• cp /work/Training/ HPM_Training/
  .
• 5. Go to /work/Training/HPM_Training/simple/

32
Lab Session for HPMRunning HPM

• 1. Compile either Fortran or C example with the
  following
• make f makefile_f (or makefile_c)
• 2. Run executable either interactive or by batch
• interactive command
• poe hpm_test.x -nodes 1
  -tasks_per_node 2 euilib ip \
• euidevice en0
• 3. Explore hpmcount summary output, looking at
  both usage and resource statistics