High-Productivity Languages for HPC: Compiler Challenges

Transcription

1 High-Productivity Languages for HPC: Compiler Challenges David Chase

2 Fortress New language Designed for productivity, high performance, abundant parallelism. Contributors Guy Steele Jan-Willem Maessen Eric Allen David Chase Sukyoung Ryu Victor Luchangco Sam Tobin-Hochstadt Yossi Lev Cheryl McCosh Joe Hallett Carl Eastlund Joao Dias Christine Flood Page 2

3 High productivity Speed of coding to scale (speed, fault tolerance) Ease of reuse Ease of debugging Ease of maintenance Portable performance (1P, CMT, SMP, NUMA, MPP) Ease of deployment (fragile dependence on DLLs?) Ease of system maintenance Larger pool of programmers Domain-specific extensions Page 3

4 High productivity features (that present compiler challenges) Garbage collection Transactional memory Fault tolerance Trustworthy compilers (Support for) cache-oblivious/work-stealing style Programming-by-contract Better human factors Page 4

5 GC, TM, FT, and the compiler GC and the compiler is pretty-well understood > Compiler can help with safepoints, barrier optimization, logging optimization, and pointer maps. > Runtime and compiler can be co-designed. > Compiler must be aware of runtime, concurrency, and memory-model issues. GC, TM, and FT are similar in many ways > Make copies of data > Monitor reads and writes > Profit from locality information > Tend to use read/write barriers, logging, and safepoints > Can we combine these? How can the optimizer help? Page 5

6 Example: Card-mark design/optimization Generational GC uses write-barriers to enforce old-young partition. Pointers from old to young must be treated specially. Traditional software write-barrier maps card X to heap addresses [X*256, X* ). A pointer written to address Y requires mark of card Y/256. Garbage collection looks for dirty cards, finds corresponding objects, and records actual oldto-young pointers. Page 6

7 Example: Card-marks and safe points Marks for multiple writes to the same address are redundant, provided no GC can intervene. (Non-concurrent) GC can only occur at safepoints; therefore, if two writes to Y are not separated by a safepoint, one card mark can be eliminated. o.f = p mark(&o.f) o.f = q mark(&o.f) o.f = p o.f = q mark(&o.f) Page 7

8 Example: Card-marks, per-object Scanning often requires access to object header; might as well scan the whole object. Objects (unlike arrays) are usually small-ish. Marks for writes to different fields can be redundant, provided no safepoint intervenes. o.f1 = p mark(&o.f1) o.f2 = q mark(&o.f2) o.f1 = p o.f2 = q mark(&o) Page 85

9 Example: Card-mark loop optimization When scanning cards, extend the range of addresses if the ending object is an array of pointers. card X maps to [X*256-64, X* ). In compiler, make writebarrier( A[i] ) be redundant with writebarrier( A[i+K] ) for 0 <= K < 8. When a loop writing into an array of objects is unrolled by 8, all but one of the write-barriers is removed (provided no safepoint intervenes) Page 9

10 Example: card-mark youth optimization. Card marking is used to record creation of pointers from old objects to young object. Newly-allocated objects are guaranteed young until a GC occurs. Until a safepoint intervenes, stores into a fresh object require no card marking. o = new O() o.f1 = p mark(&o.f1) o.f2 = q mark(&o.f2) o = new O() o.f1 = p o.f2 = q Page 10

11 Ths OS is in the way. Do we trust our compilers enough to replace it? OS threads are slow and clunky. OS traps are slow and clunky. OS virtual memory is inflexible. Do we trust our compilers enough to let them take the place of the kernel/user boundary? > We lack consensus on correctness for parallel programs. + Many correct answers; optimization may shrink that set, but that s OK. > Only an option for safe languages. What about C and C++? Page 11

12 The OS is in the way; synchronization JavaGrande Sync Method (larger is better) Total syncs per second NB Sun IBM Windows User threads OS threads Number of threads Page 12

13 The OS is in the way; wait/notifyall JavaGrande Barrier Simple (larger is better) Total barriers per second NB Sun IBM Windows User threads Number of Threads OS threads Page 13

14 Cache-oblivious computing Subdivide a problem on its largest dimension. > Good in theory, often good in practice > Automatically exploits size of caches, TLBs, working set > Good for work-stealing, work-dealing > Minimizes area of boundary between subproblems > Also N-processors oblivious But... > code generated by inlining at leaves seems to be ugly > work-stealing is spatial-locality-ignorant > can we pipeline between recursive nests? > can we map it to a cluster? Page 14

15 Programming-by-contract Said to help productivity, seems like it should. How should the optimizer use contract information? > Can we optimize contracts? > Can contracts help with library-level optimizations? > Should we only generate code for contracts that the compiler finds useful? Does the contract language allow us to say the right things? Page 15

16 Example contract for Vector class Vector { Object[] items; int size; invariant {size <= items.length} int size() { return size; } void put(int i, Object o) requires { 0 <= i < size() } { items[i] = o; } Object get(int i) requires { 0 <= i < size() } { return items[i]; } } Page 16

17 Human factors Error-reporting must be as informative as possible > Parsing > Type-inference > Exception stack traces (like Java, or Python) Error-reporting should be as early as possible Observability > Why is it slow? > Which threads have problems? How hard is it to say what I mean? > Expressive type systems can be a pain. Page 17

18 Our rotten tools; excerpt from a rant I am betting that this is the name of some C++ function, run through a one-way hash to yield a unique name. ZNSt15basic_streambufIcSt11char_traitsIcEED4Ev (Sarcastically) I assume a one-way hash, because it would just be so incredibly stupid not to fix the Unix linker for A DOZEN WHOLE YEARS to make it demangle C++ identifiers into something meaning something to the programmer.... How disgusted do I need to be, that my mail agent makes a swoosh sound IN STEREO, and my windows pour into their minaturized form, but my linker still hands me missing symbols that look like line noise. Page 18

19 High-Productivity Languages for HPC: Compiler Challenges David Chase