Partitioned Global Address Space Languages Kathy Yelick Lawrence Berkeley National Laboratory

1. Partitioned Global Address Space Languages Kathy Yelick Lawrence Berkeley National Laboratory and UC Berkeley Joint work with The Titanium Group: S. Graham, P. Hilfinger, P. Colella, D. Bonachea, K. Datta, E. Givelberg, A. Kamil, N. Mai, A. Solar, J. Su, T. Wen The Berkeley UPC Group: C. Bell, D. Bonachea, W. Chen, J. Duell, P. Hargrove, P. Husbands, C. Iancu, R. Nishtala, M. Welcome

2. The 3 P’s of Parallel Computing • Productivity • Global address space supports construction of complex shared data structures • High level constructs (e.g., multidimensional arrays) simplify programming • Performance • PGAS Languages are Faster than two-sided MPI • Some surprising hints on performance tuning • Portability • These languages are nearly ubiquitous

3. Partitioned Global Address Space • Global address space: any thread/process may directly read/write data allocated by another • Partitioned: data is designated as local (near) or global (possibly far); programmer controls layout By default: • Object heaps are shared • Program stacks are private x: 1 y: x: 5 y: x: 7 y: 0 Global address space l: l: l: g: g: g: p0 p1 pn • 3 Current languages: UPC, CAF, and Titanium • Emphasis in this talk on UPC & Titanium (based on Java)

4. PGAS Language Overview • Many common concepts, although specifics differ • Consistent with base language • Both private and shared data • int x[10]; and shared int y[10]; • Support for distributed data structures • Distributed arrays; local and global pointers/references • One-sided shared-memory communication • Simple assignment statements: x[i] = y[i]; ort = *p; • Bulk operations: memcpy in UPC, array ops in Titanium and CAF • Synchronization • Global barriers, locks, memory fences • Collective Communication, IO libraries, etc.

5. Example: Titanium Arrays • Ti Arrays created using Domains; indexed using Points: double [3d] gridA = new double [[0,0,0]:[10,10,10]]; • Eliminates some loop bound errors using foreach foreach (p in gridA.domain()) gridA[p] = gridA[p]*c + gridB[p]; • Rich domain calculus allow for slicing, subarray, transpose and other operations without data copies • Array copy operations automatically work on intersection data[neighborPos].copy(mydata); intersection (copied area) “restrict”-ed (non-ghost) cells ghost cells mydata data[neighorPos]

6. Productivity: Line Count Comparison • Comparison of NAS Parallel Benchmarks • UPC version has modest programming effort relative to C • Titanium even more compact, especially for MG, which uses multi-d arrays • Caveat: Titanium FT has user-defined Complex type and cross-language support used to call FFTW for serial 1D FFTs UPC results from Tarek El-Gazhawi et al; CAF from Chamberlain et al; Titanium joint with Kaushik Datta & Dan Bonachea

7. Case Study 1: Block-Structured AMR • Adaptive Mesh Refinement (AMR) is challenging • Irregular data accesses and control from boundaries • Mixed global/local view is useful Titanium AMR benchmarks available AMR Titanium work by Tong Wen and Philip Colella

8. Titanium AMR Entirely in Titanium Finer-grained communication No explicit pack/unpack code Automated in runtime system AMR in Titanium C++/Fortran/MPI AMR • Chombo package from LBNL • Bulk-synchronous comm: • Pack boundary data between procs 10X reduction in lines of code! *Somewhat more functionality in PDE part of Chombo code Comparable running time Work by Tong Wen and Philip Colella; Communication optimizations joint with Jimmy Su

9. Immersed Boundary Simulation in Titanium • Modeling elastic structures in an incompressible fluid. • Blood flow in the heart, blood clotting, inner ear, embryo growth, and many more • Complicated parallelization • Particle/Mesh method • “Particles” connected into materials Joint work with Ed Givelberg, Armando Solar-Lezama

10. The 3 P’s of Parallel Computing • Productivity • Global address space supports complex shared structures • High level constructs simplify programming • Performance • PGAS Languages are Faster than two-sided MPI • Better match to most HPC networks • Some surprising hints on performance tuning • Send early and often is sometimes best • Portability • These languages are nearly ubiquitous

11. PGAS Languages: High Performance Strategy for acceptance of a new language • Make it run faster than anything else Keys to high performance • Parallelism: • Scaling the number of processors • Maximize single node performance • Generate friendly code or use tuned libraries (BLAS, FFTW, etc.) • Avoid (unnecessary) communication cost • Latency, bandwidth, overhead • Berkeley UPC and Titanium use GASNet communication layer • Avoid unnecessary delays due to dependencies • Load balance; Pipeline algorithmic dependencies

12. One-Sided vs Two-Sided • A one-sided put/get message can be handled directly by a network interface with RDMA support • Avoid interrupting the CPU or storing data from CPU (preposts) • A two-sided messages needs to be matched with a receive to identify memory address to put data • Offloaded to Network Interface in networks like Quadrics • Need to download match tables to interface (from host) one-sided put message host CPU address data payload network interface two-sided message memory message id data payload

13. (up is good) Performance Advantage of One-Sided Communication: GASNet vs MPI • Opteron/InfiniBand (Jacquard at NERSC): • GASNet’s vapi-conduit and OSU MPI 0.9.5 MVAPICH • Half power point (N ½ ) differs by one order of magnitude Joint work with Paul Hargrove and Dan Bonachea

14. (down is good) GASNet: Portability and High-Performance GASNet better for latency across machines Joint work with UPC Group; GASNet design by Dan Bonachea

15. (up is good) GASNet: Portability and High-Performance GASNet at least as high (comparable) for large messages Joint work with UPC Group; GASNet design by Dan Bonachea

16. (up is good) GASNet: Portability and High-Performance GASNet excels at mid-range sizes: important for overlap Joint work with UPC Group; GASNet design by Dan Bonachea

17. Case Study 2: NAS FT • Performance of Exchange (Alltoall) is critical • 1D FFTs in each dimension, 3 phases • Transpose after first 2 for locality • Bisection bandwidth-limited • Problem as #procs grows • Three approaches: • Exchange: • wait for 2nd dim FFTs to finish, send 1 message per processor pair • Slab: • wait for chunk of rows destined for 1 proc, send when ready • Pencil: • send each row as it completes Joint work with Chris Bell, Rajesh Nishtala, Dan Bonachea

18. Overlapping Communication • Goal: make use of “all the wires all the time” • Schedule communication to avoid network backup • Trade-off: overhead vs. overlap • Exchange has fewest messages, less message overhead • Slabs and pencils have more overlap; pencils the most • Example: Class D problem on 256 Processors Joint work with Chris Bell, Rajesh Nishtala, Dan Bonachea

19. NAS FT Variants Performance Summary .5 Tflops • Slab is always best for MPI; small message cost too high • Pencil is always best for UPC; more overlap Joint work with Chris Bell, Rajesh Nishtala, Dan Bonachea

20. Case Study 2: LU Factorization • Direct methods have complicated dependencies • Especially with pivoting (unpredictable communication) • Especially for sparse matrices (dependence graph with holes) • LU Factorization in UPC • Use overlap ideas and multithreading to mask latency • Multithreaded: UPC threads + user threads + threaded BLAS • Panel factorization: Including pivoting • Update to a block of U • Trailing submatrix updates • Status: • Dense LU done: HPL-compliant • Sparse version underway Joint work with Parry Husbands

21. UPC HPL Performance • Comparison to ScaLAPACK on an Altix, a 2 x 4 process grid • ScaLAPACK (block size 64) 25.25 GFlop/s (tried several block sizes) • UPC LU (block size 256) - 33.60 GFlop/s, (block size 64) - 26.47 GFlop/s • n = 32000 on a 4x4 process grid • ScaLAPACK - 43.34 GFlop/s (block size = 64) • UPC - 70.26 Gflop/s (block size = 200) • MPI HPL numbers from HPCC database • Large scaling: • 2.2 TFlops on 512p, • 4.4 TFlops on 1024p (Thunder) Joint work with Parry Husbands

22. The 3 P’s of Parallel Computing • Productivity • Global address space supports complex shared structures • High level constructs simplify programming • Performance • PGAS Languages are Faster than two-sided MPI • Some surprising hints on performance tuning • Portability • These languages are nearly ubiquitous • Source-to-source translators are key • Combined with portable communication layer • Specialized compilers are useful in some cases

23. Portability of Titanium and UPC • Titanium and the Berkeley UPC translator use a similar model • Source-to-source translator (generate ISO C) • Runtime layer implements global pointers, etc • Common communication layer (GASNet) • Both run on most PCs, SMPs, clusters & supercomputers • Support Operating Systems: • Linux, FreeBSD, Tru64, AIX, IRIX, HPUX, Solaris, Cygwin, MacOSX, Unicos, SuperUX • UPC translator somewhat less portable: we provide a http-based compile server • Supported CPUs: • x86, Itanium, Alpha, Sparc, PowerPC, PA-RISC, Opteron • GASNet communication: • Myrinet GM, Quadrics Elan, Mellanox Infiniband VAPI, IBM LAPI, Cray X1, SGI Altix, Cray/SGI SHMEM, and (for portability) MPI and UDP • Specific supercomputer platforms: • HP AlphaServer, Cray X1, IBM SP, NEC SX-6, Cluster X (Big Mac), SGI Altix 3000 • Underway: Cray XT3, BG/L (both run over MPI) • Can be mixed with MPI, C/C++, Fortran Also used by gcc/upc Joint work with Titanium and UPC groups

24. Portability of PGAS Languages Other compilers also exist for PGAS Languages • UPC • Gcc/UPC by Intrepid: runs on GASNet • HP UPC for AlphaServers, clusters, … • MTU UPC uses HP compiler on MPI (source to source) • Cray UPC • Co-Array Fortran: • Cray CAF Compiler: X1, X1E • Rice CAF Compiler (on ARMCI or GASNet), John Mellor-Crummey • Source to source • Processors: Pentium, Itanium2, Alpha, MIPS • Networks: Myrinet, Quadrics, Altix, Origin, Ethernet • OS: Linux32 RedHat, IRIS, Tru64 NB: source-to-source requires cooperation by backend compilers

25. Summary • PGAS languages offer performance advantages • Good match to RDMA support in networks • Smaller messages may be faster: • make better use of network: postpone bisection bandwidth pain • can also prevent cache thrashing for packing • PGAS languages offer productivity advantage • Order of magnitude in line counts for grid-based code in Titanium • Push decisions about packing/not into runtime for portability (advantage of language with translator vs. library approach) • Source-to-source translation • The way to ubiquity • Complement highly tuned machine-specific compilers

26. End of Slides

27. Productizing BUPC • Recent Berkeley UPC release • Support full 1.2 language spec • Supports collectives (tuning ongoing); memory model compliance • Supports UPC I/O (naïve reference implementation) • Large effort in quality assurance and robustness • Test suite: 600+ tests run nightly on 20+ platform configs • Tests correct compilation & execution of UPC and GASNet • >30,000 UPC compilations and >20,000 UPC test runs per night • Online reporting of results & hookup with bug database • Test suite infrastructure extended to support any UPC compiler • now running nightly with GCC/UPC + UPCR • also support HP-UPC, Cray UPC, … • Online bug reporting database • Over >1100 reports since Jan 03 • > 90% fixed (excl. enhancement requests)

28. Benchmarking • Next few UPC and MPI application benchmarks use the following systems • Myrinet: Myrinet 2000 PCI64B, P4-Xeon 2.2GHz • InfiniBand: IB Mellanox Cougar 4X HCA, Opteron 2.2GHz • Elan3: Quadrics QsNet1, Alpha 1GHz • Elan4: Quadrics QsNet2, Itanium2 1.4GHz

29. PGAS Languages: Key to High Performance One way to gain acceptance of a new language • Make it run faster than anything else Keys to high performance • Parallelism: • Scaling the number of processors • Maximize single node performance • Generate friendly code or use tuned libraries (BLAS, FFTW, etc.) • Avoid (unnecessary) communication cost • Latency, bandwidth, overhead • Avoid unnecessary delays due to dependencies • Load balance • Pipeline algorithmic dependencies

30. Hardware Latency • Network latency is not expected to improve significantly • Overlapping communication automatically (Chen) • Overlapping manually in the UPC applications (Husbands, Welcome, Bell, Nishtala) • Language support for overlap (Bonachea)

31. Effective Latency Communication wait time from other factors • Algorithmic dependencies • Use finer-grained parallelism, pipeline tasks (Husbands) • Communication bandwidth bottleneck • Message time is: Latency + 1/Bandwidth * Size • Too much aggregation hurts: wait for bandwidth term • De-aggregation optimization: automatic (Iancu); • Bisection bandwidth bottlenecks • Spread communication throughout the computation (Bell)

32. Fine-grained UPC vs. Bulk-Synch MPI • How to waste money on supercomputers • Pack all communication into single message (spend memory bandwidth) • Save all communication until the last one is ready (add effective latency) • Send all at once (spend bisection bandwidth) • Or, to use what you have efficiently: • Avoid long wait times: send early and often • Use “all the wires, all the time” • This requires having low overhead!

33. What You Won’t Hear Much About • Compiler/runtime/gasnet bug fixes, performance tuning, testing, … • >13,000 e-mail messages regarding cvs checkins • Nightly regression testing • 25 platforms, 3 compilers (head, opt-branch, gcc-upc), • Bug reporting • 1177 bug reports, 1027 fixed • Release scheduled for later this summer • Beta is available • Process significantly streamlined

34. Take-Home Messages • Titanium offers tremendous gains in productivity • High level domain-specific array abstractions • Titanium is being used for real applications • Not just toy problems • Titanium and UPC are both highly portable • Run on essentially any machine • Rigorously tested and supported • PGAS Languages are Faster than two-sided MPI • Better match to most HPC networks • Berkeley UPC and Titanium benchmarks • Designed from scratch with one-side PGAS model • Focus on 2 scalability challenges: AMR and Sparse LU

35. Titanium Background • Based on Java, a cleaner C++ • Classes, automatic memory management, etc. • Compiled to C and then machine code, no JVM • Same parallelism model at UPC and CAF • SPMD parallelism • Dynamic Java threads are not supported • Optimizing compiler • Analyzes global synchronization • Optimizes pointers, communication, memory

36. Do these Features Yield Productivity? Joint work with Kaushik Datta, Dan Bonachea

37. GASNet/X1 Performance single word get single word put GASNet/X1 improves small message performance over shmem and MPI Leverages global pointers on X1 Highlights advantage of languages vs. library approach Joint work with Christian Bell, Wei Chen and Dan Bonachea

38. High Level Optimizations in Titanium • Irregular communication can be expensive • “Best” strategy differs by data size/distribution and machine parameters • E.g., packing, sending bounding boxes, fine-grained are • Use of runtime optimizations • Inspector-executor • Performance on Sparse MatVec Mult • Results: best strategy differs within the machine on a single matrix (~ 50% better) Speedup relative to MPI code (Aztec library) Average and maximum speedup of the Titanium version relative to the Aztec version on 1 to 16 processors Joint work with Jimmy Su

39. Source to Source Strategy • Source-to-source translation strategy • Tremendous portability advantage • Still can perform significant optimizations • Relies on high quality back-end compilers and some coaxing in code generation Use of “restrict” pointers in C Understand Multi-D array indexing (Intel/Itanium issue) Support for pragmas like IVDEP Robust vectorizators (X1, SSE, NEC,…) 48x On machines with integrated shared memory hardware need access to shared memory operations Joint work with Jimmy Su