HP Caliper

1
HP Caliper

• Eric Gouriou
• September 2003

2
Todays Agenda

• Intended audience of this presentation
• What is HP Caliper ?
• Measurements
• Usage
• Caliper cheat sheet
• Limitations
• Future plans
• More information on Caliper, DSPP
• Questions

3
Intended Audience

• Developers
• Tuning experts
• Engineers porting to HP-UX IPF
• Anyone interested in performance

4
What is HP Caliper ?

• Performance analysis improvement tool
• Dynamic performance measurement tool forC / C
  / Fortran / assembly applications
• Data collection vehicle for compiler-feedback/PBO
• Works on all programs as is(32/64bits, debug or
  optimized, stripped, etc.)
• Multiple measurements via a unified interface
• Provides insights thanks to
• Itanium Performance Monitoring Unit (PMU)
• Dynamic binary instrumentation

5
Key Features

• Default measurement configurations, configurable
• Selective process module measurements
• Text HTML reports
• Performance datafile
• Three measurement models
• start application under Caliper
• attach to running process
• auto-invocation

6
Measurements

• PMU event counts
• Total count of selected hardware events per
  process
• Negligible overhead
• Default set of events, can be overriden
• 400 events described in Itanium2 documentation
• Non-default use for advanced users
• ------------------------------------------------
• Counter Priv. Mask Count
• -------------------------------------------
• IA64_INST_RETIRED 8 (USER) 3414917
• NOPS_RETIRED 8 (USER) 684477
• CPU_CYCLES 8 (USER) 1899187
• BACK_END_BUBBLE_ALL 8 (USER) 810631
• -------------------------------------------
• of Cycles lost due to stalls (lower is
  better)
• 42.68 100 BACK_END_BUBBLE_ALL / CPU_CYCLES
• Effective CPI (lower is better)

7
Measurements (contd)

• Histograms from samples of PMU data
• Allows identification of hotspots
• Module summary, function summary, function
  details, selected global counts and derived
  metrics
• Flat profile, Cache / TLB / Branch Prediction /
  ALAT
• Per-thread data available
• Very low overhead
• Function Summary
• -------------------------------------------------
  ------------------------
• Total Cumulat
  
• IP of IP
  
• Samples Total Samples Function
  File
• -------------------------------------------------
  ------------------------
• 9.71 9.71 1286
  livermoreinit livermore.c
• 6.03 15.75 799
  livermoremain livermore.c
• 4.07 19.82 539
  libc.so.1T_19_3c30_cl___doprnt_main doprnt.c
• 0.60 20.42 79
  libc.so.1_f80_to_dec bindec.c
• 0.45 20.86 59
  libc.so.1getenv getenv.c
• ...

8
Measurements (contd)

• Traces of PMU samples
• Provides full details for each sample
• Low overhead but high volume of data
• Customize configuration file for relevant data
• -------------------------------------------------
  --------------------------------------
• -----------------------DCache
  Miss---------------------- ------IP Samples------
• Sample AddrSlot Data Bundle
  Bundle Address
• Number (modulefunction) Runtime
  Address Latency (modulefunction)
• -------------------------------------------------
  --------------------------------------
• 1 0x3eda00
  0x200000007979f700 5 0x502f0
  
• (dld.soMM_malloc)
  (dld.soBU_grow)
• 2 0x211c00
  0x200000007950b200 26 0x212a0
  
• (dld.soLE_finish_create)
  (dld.soLE_finish_create)
• 3 0x37bf00
  0x20000000795297b8 172 0x37c40
  
• (dld.soR_apply_eplt_relocs)
  (dld.soR_apply_eplt_relocs)
• ...

9
Measurements (contd)

• Source-level event counts
• Function call counts, arc counts
• High overhead, precise counts
• Done via dynamic binary instrumentation
• Function Count Details
• -------------------------------------------------
  ---------
• Total Function
  File Line
• -------------------------------------------------
  ---------
• 150 livermoreabs
  livermore.c 405
• 104 libc.so.1__milli_memset
• 92 libc.so.1__milli_memmove
• ...
• Arc Counts
• -------------------------------------------------
  -------------------------------------------
• Total Taken Taken Source Address
  Source Function File Line,Col
• Target Address
  Target Function File Line,Col
• -------------------------------------------------
  -------------------------------------------
• 28672 28616 99 0x4005e702
  livermoreinit livermore.c376,7

10
Measurements (contd)

• Call graph profile (gprof-like)
• Flat profile and call graph
• High overhead
• Hybrid of exact counts and PMU sampling
• Call Graph
• -------------------------------------------------
  -------------
• De- Called/Total
  Parents
• Index Time Self scen- CalledSelf
  Name Index
• dants Called/Total
  Children
• -------------------------------------------------
  -------------
• 0.00 0.00 1/1
  ROOT 1
• 2 25.1 0.00 0.00 1
  livermoremain 2
• 0.00 0.00 150/150
  livermoreabs 52
• 0.00 0.00 30/30
  livermoreclock 45
• 0.00 0.00 14/14
  livermoreinit 3
• 0.00 0.00 18/18
  libc.so.1printf 4
• -------------------------------------------------
  -------------
• 0.00 0.00 14/14
  livermoremain 2

11
Measurements (contd)

• PBO profile gathering configuration
• Auto-invoked when compiling using Oprofilecolle
  ct (I deprecated)
• Data used to improve compiler optimizations Opro
  fileuse (P deprecated)
• Can be done manually (caliper pbo ...), however
  not recommended, sub-optimal
• Variable overhead
• gt cc Oprofilecollect -o livermore livermore.c
• gt ./livermore
• ...
• gt ls flow
• flow.data flow.log
• gt cc Oprofileuse O3 -o livermore livermore.c

12
Usage

• Typical command line
• caliper config_file caliper_options program
  program_args
• Example
• gt caliper fprof --processall cc -o livermore
  livermore.c
• Configuration files
• packaged ones
• copy/modify
• command-line overrides

13
Usage (contd)

• Type Configuration Files Comments
• --------------------------------------------------
  --------------------------------------------
• Histograms fprof, reduced samples,
• dicache_miss very low impact
• ditlb_miss
• branch_prediction
• alat_miss
• Call graph cgprof sampled exact, high
  impact
• Sampled details pmu_trace large data volume
• Total HW event counts total_cpu exact totals,
  no impact
• Exact source-level arc_count, exact
  details,event counts func_count, high impact
• Compiler feedback pbo black box

14
Caliper Cheat Sheet

• Where should I start ?
• Global view
• fprof, both for profile and per-process derived
  metrics
• cgprof, caller/callee, check for surprises
• dcache_miss, use latency threshold to show
  expensive misses
• Drill-down
• Restrict processes, libraries, functions measured
• What is missing for a global view ?
• System-wide measurements
• Multiplexed global counts (vs. many total_cpu
  runs)

15
Caliper Cheat Sheet (contd)

• Tuning the data collection parameters
• Multi-process application ?Check process tree
  output, select processes using --process
• Multi-threaded application ?Check --threadsall (
  per-thread histograms)versus --threadssum-all (d
  efault, aggregated data)
• Libraries of interest or out of your control ?
• Use --module-include / --module-exclude
• Functions of interest ?
• Check --user-regionsrum/sum and/or triggered
  samples

16
Caliper Cheat Sheet (contd)

• Better reports
• Use HTML output (--html), text is the default
• Use datafiles
• Allow multiple reports for a single run
• Faster collection in multi-process runs
• Check source-level reporting
• --report-detailsstatement
• Vary amount of details generated

17
Caliper Cheat Sheet (contd)

• PBO
• Performance for free for some applications
  (almost)
• Use Oprofilecollect/use
• caliper pbo works on O1 binaries but isnt
  recommended
• Can use chatr I enable to enable
  auto-invocation
• Trade-offs for large multi-process
  applications,1 vs. many Caliper

18
Limitations

• Application characteristics
• no dynamic library reload
• Measurement control
• pbo profile collection requires O1
  binary(automatic when using Oprofilecollect)
• HW limits the measurements possible per run
• per-thread data limited to histograms
• Other
• emulated PA binaries are not measured
• minimal dynamic code support
• limited gcc/g support
• setuid binaries require workaround
• limited support for MPAS binaries

19
Future Plans

• PMU Measurements
• multiplexed PMU runs
• richer derived metrics
• system-wide measurements
• kernel profiles
• PBO
• PMU cache data collected for PBO
• Data Files
• aggregation
• merging
• diffing

20
Future Plans (contd)

• Usability
• Graphs w/ html reports
• Reports on demand
• Function context
• Call stacks
• Remove limitations
• Detach for runs involving instrumentation
• MPAS applications
• Library load/unload
• Dynamically generated code

21
More Information

• The Caliper web page is on the DSPP website
• lthttp//www.hp.com/go/hpcalipergt
• Documentation / Support / Downloads
• The Caliper mailing lists
• Majordomo lists ltmajordomo_at_cxx.cup.hp.comgt
• For product announcements
• ltcaliper-announcegt
• For announcements and user forum
• ltcalipergt

22
DSPP Tools Resources for Itanium 2 Set You Up
for Success

• Software
• development environments, compilers, operating
  systems, installation/configuration tools,
  performance tools and more
• Technical documentation
• white papers, tutorials, references documents and
  manuals, FAQs, known problems, sample code, etc.
  
• Training and Education
• online and classroom training

23
More DSPP Tools Resources

• Community
• Itanium forums, source code repository, document
  sharing and mailing lists
• Equipment
• rentals and purchase discounts
• Partner Resources
• News Events

D S P P
24
Where to go

• Start with the Itanium web site for DSPP
  partners
• http//www.hp.com/go/dspp_itanium
• Contact points for additional information,
  general support,
• equipment, localization resources and more
• Americas spp_at_cup.hp.com
• telephone [phone]
• Europe dspp.emea_at_hp.com
• telephone [phone]
• Asia-Pac hpdev.support_at_hp.com or go to
  www.hp.com/go/dspp for local country
  contacts

25
Quote slide
Questions?
26
(No Transcript)