Robert Fowler John Mellor-Crummey Nathan Tallent Gabriel Marin Department of Computer Science

1. Performance Tuning on Modern Systems: Tools and Experiences Robert Fowler John Mellor-Crummey Nathan Tallent Gabriel Marin Department of Computer Science Rice University http://hipersoft.cs.rice.edu/hpctoolkit/

2. Thought for the Day 'The Hitchiker's Guide to the Galaxy, in a moment of reasoned lucidity which is almost unique among its current tally of five million, nine hundred and seventy-three thousand, five hundred and nine pages, says of the Sirius Cybernetics Corporation products that “it is very easy to be blinded to the essential uselessness of them by the sense of achievement you get from getting them to work at all. In other words - and this is the rock-solid principle on which the whole of the Corporation's galaxywide success is founded -- their fundamental design flaws are completely hidden by their superficial design flaws.” (Douglas Adams, "So Long, and Thanks for all the Fish")

3. Background What we (Rice Parallel Compilers Group) do. • Code optimization. • Aggressive (mostly source-to-source) compiler optimization. • Hand application of transformations. • Try out transformations by hand first. • Work on real codes with algorithm, library, and application developers. • (Deployment time for compiler research is measured in years.). Why we spend a lot of time (too much?) analyzing executions. Why? • It is hard to understand deeply pipelined, out of order, superscalar processors with non-blocking caches and deep memory hierarchies. • Aggressive ( -O4) and/or idiosyncratic vendor compilers make life harder. What we did. • Built tools to meet our needs w.r.t. the run/analyze/tune cycle..

4. LMBENCH NUMBERS

5. Streams numbers for native compiler and gcc

6. Memory Bandwidth: Fundamental Performance Limit • Example: simple 2-d relaxation kernel S:r(i,j) = a * x(i,j) + b * (x(i-1,j)+x(i+1,j)+x(i,j-1)+x(i,j+1)) • Each execution of S performs 6 ops, uses 5 x values. (writes 1 result) (not counting index computations or scalars kept in registers) • Assume x is double and always comes from main memory • S consumes 40 bytes of memory read bandwidth, 6.7 bytes/op. • 1GB/sec memory  150MFLOPS upper bound, 25M updates/sec • But in each outer loop, each x value is used exactly 5 times, 6 if r  a, so if a clever programmer/compiler gets perfect reuse in registers and cache: • S consumes 5*(1/5)*8 = 8 bytes of memory read bandwidth, 1.3bytes/op. • 1GB/sec memory  750MFLOPS upper bound, 125M updates/sec • First latency has to be hidden • Load scheduling • Prefetching • Software pipelining • There are lots of techniques for getting this kind of reuse • Tiling -- good for cache reuse • Unroll and jam – both cache and register reuse

7. Memory Bandwidth: Part 2 Example: realistic 2-d relaxation kernel S:r(i,j) = c(i,j)*x(i,j) + c(i-1,j)*x(i-1,j) + c(i+1,j)*x(i+1,j) + c(i,j-1)*x(i,j-1) + c(i,j+1)*x(i,j+1) • Each execution of S performs 9 ops, uses 5 x and 5 c values,(not counting index computations or scalars kept in registers) • Assume x is double and always comes from main memory • S consumes 80 bytes of memory read bandwidth, 8.9 bytes/op. • 1GB/sec memory  112MFLOPS upper bound, 12.5 M updates/sec • In each outer loop, each x value is used exactly 5 times, 6 if r  a.But each c value is used only once (twice for i≠j if c is symmetric). • S consumes (5 + 1*1/5)*8 = 48 bytes of memory read bandwidth, 5.33bytes/op. • 1GB/sec memory  187MFLOPS upper bound, 20.8M updates/sec. • Solutions • “Superficial stuff” • hide latency and get all the reuse you can. • “Fundamental stuff” • Aggressive algorithms to increase reuse (Time skewing, etc) • Other algorithms? • Who cares if the flop rate is low if the solution is better, faster?

8. Performance Analysis and Tuning • Increasingly necessary • Gap between typical and peak performance is growing. • Increasingly hard • Complex architectures are harder to program effectively • complex processors • VLIW • deeply pipelined, out of order, superscalar • complex memory hierarchy • non-blocking, multi-level caches • TLB • Modern scientific applications pose challenges for tools • multi-lingual programs • many source files • complex build process • external libraries in binary-only form • Tools have superficial design problems and aren’t used.

9. A 32+ Year History of Disappointment Historical origins • Johnson & Johnston (March 1970) – • IBM 360/91 at SLAC. • Profiling on OS360 MVT using timer in control process. • Knuth (circa 1973) characterization of Fortran programs– • The 90/10 rule (or was it 80/20?) • Write a clear, correct initial version • Use profiling tools to identify the hot spots • Focus on the hot spots. Since then • Three decades of academic and industrial tools • Sporadic advocacy in CS and Scientific Computing courses But • Compelling success stories are rare. • Tools are underutilized outside a few small communities. A 10% hotspot in a 1M line application could be 100K linesscattered throughout thep rogram.

10. Problems with Existing Performance Tools Most tools are hard to use on big, complex, modular applications. Multiple libraries, libraries, runtime linking, multiple languages Tools (feel like they) are designed to evaluate architectural choices or OS design, rather than for application tuning. User Interface Issues GUI problems Platform specific, both for instrumentation and analysis workstations Non-portable, non-collaborative visualization Single-metric displays don’t capture underlying problems Need to aggregate data by block, loop, and user-defined scopes Failure to make compelling cases. Users need to dig to find problems and solutions. Hard to convince management, application developers with a fat stack of printouts and hand-generated notes.

11. More Problems Language- or compiler-based tools can’t handle big applications. They can’t handle multi-language, multi-module (library) programs. Even vendor compilers are challenged (correctness and optimization) by big programs. Insufficient analytic power in tools makes analysis difficult/tedious Any one performance metric produces myopic view. Some metrics are causes. Some are effects. Relationship? Ex. – Cache misses problematic only if latency isn’t hidden. Ex. – Instruction balance: loads/stores, FLOPS, integer. Ex. – Fusion of compiler analysis, simulation, and counter profiles. Labor intensive tasks are difficult and/or tedious. (Re-)inserting explicit instrumentation and/or recompiling. Manual correlation of data from multiple sources with source code. Aggregating per line data into larger scopes. Computing synthetic measures, e.g. loads/flop. Tune/reanalyze cycle slowed or prevented by manual tasks. Manual (re-)insertion of instrumentation points. Manual (re-)computation of derived metrics. Manual (re-)correlation of performance data with source constructs.

12. HPCToolkit Goals • Support large, multi-lingual applications • a mix of of Fortran, C, C++ • external libraries • thousands of procedures • hundreds of thousands of lines • we must avoid • manual instrumentation • significantly altering the build process • frequent recompilation • Multi-platform • Scalable data collection • Analyze both serial and parallel codes • Effective presentation of analysis results • intuitive enough for physicists and engineers to use • detailed enough to meet the needs of compiler writers

13. binary object code compilation linking source correlation binary analysis profile execution program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow application source Drive this with scripts. Call scripts in Makefiles. On parallel systems, integrate scripts with batch system.

14. HPCToolkit Workflow application source binary object code compilation • launch unmodified, optimized application binaries • collect statistical profiles of events of interest linking source correlation binary analysis profile execution program structure hyperlinked database performance profile interpret profile hpcviewer

15. application source binary object code compilation linking source correlation profile execution binary analysis program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow • decode instructions and combine with profile data

16. application source binary object code compilation linking source correlation profile execution binary analysis program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow • extract loop nesting information from executables

17. application source binary object code compilation linking source correlation profile execution binary analysis program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow • synthesize new metrics by combining metrics • relate metrics, structure, and program source

18. application source binary object code compilation linking source correlation profile execution binary analysis program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow • support top-down analysis with interactive viewer • analyze results anytime, anywhere

19. HPCToolkit Workflow application source binary object code compilation linking source correlation binary analysis profile execution program structure hyperlinked database performance profile interpret profile hpcviewer

20. Data Collection Support analysis of unmodified, optimized binaries • Inserting code to start, stop and read counters has many drawbacks, so don’t do it! • Nested measurements skew results • Instrumentation points interfere with optimization. • Use hardware performance monitoring to collect statistical profiles of events of interest • Different platforms have different capabilities • event-based counters: MIPS, IA64, AMD64, IA32 • ProfileMe instruction tracing: Alpha • Different capabilities require different approaches

21. Data Collection Tools Goal: limit development to essentials only • MIPS-IRIX: • ssrun + prof ptran • Alpha-Tru64: • uprofile + prof ptran • DCPI/ProfileMe xprof • IA64-Linux and IA32-Linux • papirun/papiprof

22. papirun/papiprof • PAPI: Performance API • interface to hardware performance monitors • supports many platforms • papirun: open source equivalent of SGI’s ‘ssrun’ • sample-based profiling of an execution • preload monitoring library before launching application • inspect load map to set up sampling for all load modules • record PC samples for each module along with load map • Linux IA64 and IA32 • papiprof: ‘prof’-like tool • based on Curtis Janssen’s vprof • uses GNU binutils to perform PC  source mapping • output styles • XML for use with hpcview • plain text

23. DCPI and ProfileMe • Alpha ProfileMe • EV67+ records info about an instruction as it executes • mispredicted branches, memory access replay traps • more accurate attribution of events • DCPI: (Digital) Continuous Profiling Infrastructure • sample processor counters and instructions continuously during execution of all code • all programs • shared libraries • operating system • support both on-line and off-line data analysis • to date, we use only off-line analysis

24. application source binary object code compilation linking source correlation profile execution binary analysis program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow

25. Metric Synthesis with xprof (Alpha) Interpret DCPI/ProfileMe samples into useful metrics • Transform low-level data to higher-level metrics • DCPI/ProfileMe information associated with PC values • project ProfileMe data into useful equivalence classes • decode instruction type info in application binary at each PC • FLOP • memory operation • integer operation • fuse the two kinds of information • Retired instruction sample points + instruction type  • retired FLOPs • retired integer operations • retired memory operations • Map back to source code like papiprof

26. application source binary object code compilation linking source correlation profile execution binary analysis program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow

27. Program Structure Recovery with bloop • Parse instructions in an executable using GNU binutils • Analyze branches to identify basic blocks • Construct control flow graph using branch target analysis • be careful with machine conventions and delay slots! • Use interval analysis to identify natural loop nests • Map machine instructions to source lines with symbol table • dependent on accurate debugging information! • Normalize output to recover source-level view Platforms: Alpha+Tru64, MIPS+IRIX, Linux+IA64, Linux+IA32, Solaris+SPARC

28. Sample Flowgraph from an Executable Loop nesting structure • blue: outermost level • red: loop level 1 • green loop level 2 Observation optimization complicates program structure!

29. Normalizing Program Structure Coalesce duplicate lines (1) if duplicate lines appear in different loops • find least common ancestor in scope tree; merge corresponding loops along the paths to each of the duplicates • purpose: re-rolls loops that have been split (2) if duplicate lines appear at multiple levels in a loop nest • discard all but the innermost instance • purpose: handles loop-invariant code motion apply (1) and (2) repeatedly until a fixed point is reached Constraint:each source line must appear at most once

30. Load Module File Procedure Loop Statement Recovered Program Structure <LM n="/apps/smg98/test/smg98"> ... <F n="/apps/smg98/struct_linear_solvers/smg_relax.c"> <P n="hypre_SMGRelaxFreeARem"> <L b="146" e="146"> <S b="146" e="146"/> </L> </P> <P n="hypre_SMGRelax"> <L b="297" e="328"> <S b="297" e="297"/> <L b="301" e="328"> <S b="301" e="301"/> <L b="318" e="325"> <S b="318" e="325"/> </L> <S b="328" e="328"/> </L> <S b="302" e="302"/> </L> </P> ... </F> </PGM>

31. application source binary object code compilation linking source correlation profile execution binary analysis program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow

32. Data Correlation • Problem • any one performance measure provides a myopic view • some measure potential causes (e.g. cache misses) • some measure effects (e.g. cycles) • cache misses not always a problem • event counter attribution is inaccurate for out-of-order processors • Approaches • multiple metrics for each program line • computed metrics, e.g. cycles - FLOPS • eliminate mental arithmetic • serve as a key for sorting • hierarchical structure • line level attribution errors give good loop-level information

33. application source binary object code compilation linking source correlation profile execution binary analysis program structure hyperlinked database performance profile interpret profile hpcviewer HPCToolkit Workflow

34. HPCViewerScreenshot Annotated Source View Navigation Metrics

35. Flattening for Top Down Analysis • Problem • strict hierarchical view of a program is too rigid • want to compare program components at the same level as peers • Solution • enable a scope’s descendants to be flattened to compare their children as peers Current scope flatten unflatten

36. Some Uses for HPCToolkit • Identifying unproductive work • where is the program spending its time not performing FLOPS • Memory hierarchy issues • bandwidth utilization: misses x line size/cycles • exposed latency: ideal vs. measured • Cross architecture or compiler comparisons • what program features cause performance differences? • Gap between peak and observed performance • loop balance vs. machine balance? • Evaluating load balance in a parallelized code • how do profiles for different processes compare

37. Assessment of HPCToolkit Functionality • Top down analysis focuses attention where it belongs • sorted views put the important things first • Integrated browsing interface facilitates exploration • rich network of connections makes navigation simple • Hierarchical, loop-level reporting facilitates analysis • more sensible view when statement-level data is imprecise • Binary analysis handles multi-lingual applications and libraries • succeeds where language and compiler based tools can’t • Sample-based profiling, aggregation and derived metrics • reduce manual effort in analysis and tuning cycle • Multiple metrics provide a better picture of performance • Multi-platform data collection • Platform independent analysis tool

38. What’s Next? Research • collect and present dynamic content • what path gets us to expensive computations? • accurate call-graph profiling of unmodified executables • analysis and presentation of dynamic content • communication in parallel programs • statistical clustering for analyzing large-scale parallelism • performance diagnosis: why rather than what Development • harden toolchain • new platforms: Opteron and PowerPC(Apple) • data collection with oprofile on Linux • distributed and collaborative extensions

39. Contacts HPCView tools: http://www.cs.rice.edu/~dsystem/hpcview/ Group: hpc@cs.rice.edu Rob: rjf@rice.edu John: johnmc@rice.edu

40. RISC  “High ILP” processors: What went wrong? • The “RISC revolution” was predicated on the assumption that aggressive compiler technology was mature and would be able to generate optimized code that can efficiently utilize these machines. • “ILP features” (superscalar, VLIW, non-blocking caches) increase the reliance on effective compilation. But • In practice, many very aggressive optimizations fail to be applied to real programs. Implementation bugs? Incorrect assumptions? Aspects of the apps? • A big scientific code that gets 5% of the peak flop rate considered to be doing well • Some big scientific codes get <0.05 FLOP/cycle. • Even well compiled programs are bandwidth bound. • Why not just use the cheapest Athalon compatible with the fastest chipset/memory combination?