Unified Parallel C UPC and the Berkeley UPC Compiler

1
Unified Parallel C (UPC) and the Berkeley UPC
Compiler
Wei Chen the Berkeley UPC Group 3/11/07
2
Parallel Programming

• Most parallel programs are written using either
• Message passing with a SPMD model
• Usually for scientific applications with
  C/Fortran
• Scales easily user controlled data layout
• Hard to use send/receive matching, message
  packing/unpacking
• Shared memory with OpenMP/pthreads/Java
• Usually for non-scientific applications
• Easier to program direct reads and writes to
  shared data
• Hard to scale (mostly) limited to SMPs, no
  concept of locality
• PGAS an alternative hybrid model

3
Partitioned Global Address Space

• PGAS model uses global address space abstraction
• Shared memory is partitioned by processors
• User controlled data layout (global pointers and
  distributed arrays)
• One-sided communication
• Use RDMA support for reads/writes of shared
  variables
• Much faster than message passing for small/medium
  size messages
• Hybrid model works for both SMPs and clusters
• Languages Titanium, Co-Array Fortran, UPC

X0
X1
XP
Shared
Global address space
ptr
ptr
ptr
Private
4
Unified Parallel C

• A SPMD parallel extension of C
• PGAS add shared qualifier to type system
• Several kinds of shared array distributions
• Fine-grained and bulk communication
• Commercial compilers with Cray/HP/IBM
• Open source compilers with Berkeley UPC

Vector Addition in UPC
define N 100THREADSshared int v1N, v2N,
sumN //cyclic layoutvoid main() for(int
i0 iltN i) if (MYTHREAD iTHREADS)
//SPMD sumiv1iv2i
5
Overview of the Berkeley UPC Compiler
Two Goals Portability and High-Performance
Lower UPC code into ISO C code
Translator
UPC Code
Shared Memory Management and pointer operations
Platform- independent
Translator Generated C Code
Berkeley UPC Runtime System
Network- independent
Compiler- independent
GASNet Communication System
Language- independent
Network Hardware
Uniform get/put interface for underlying networks
6
UPC to C Translator

• Based on Open64
• Extend with shared type
• Reuse analysis framework
• Add UPC specific optimizations
• Portable translation
• High level IR
• Config file for platform dependent information
• Reinclude library headers
• Convert shared memory operations into runtime
  calls

Preprocessed UPC Source
Parsing
WHIRL with shared types
Optimizer
Optimized WHIRL
Lowering
WHIRL with runtime calls
Lowering
WHIRL2C
Backend C compiler
ISO C code
7
Optimization framework

• Combination of language/compiler/runtime support
• Transparent to the user
• Performance portable
• Short term goal effective on different cluster
  networks.
• Long term goal code designed for SMP get good
  performance on clusters

Optimize regular array accesses
Optimize irregular pointer accesses
Nonblocking bulk communication
p-gtx-gty
upc_memget(dst, src, size)
Aijk
Loop framework for message vectorization, strip
mining
PRE framework with split-phase access and
coalescing
Runtime framework for communication overlap
8
Application Performance LU Decomposition

• UPC performance comparable to MPI/HPL(Linpack)
  with lt ½ the code size
• Uses light-weight multi-threading atop SPMD ?
  latency tolerant
• Highly adaptable to different problem and machine
  sizes

9
Application Performance 3D FFT
MFLOPS / Proc
up is good

• One-sided UPC approach sends more, smaller
  messages
• Same total volume of data, but send earlier and
  more often
• Aggressively overlaps the transpose with the 2nd
  1-D FFT
• Same approach is less effective in MPI due to
  higher per-message cost
• Consistently outperforms MPI-based
  implementations by as much as 2X

10
Current Status

• Public release v2.4 in November 2006
• Fully compliant with UPC 1.2 specification
• Communication optimizations
• Extensions for performance and programmability
• Support from laptops to supercomputers
• OS UNIX (Linux, BSD, AIX, Solaris, etc), Mac,
  Cygwin
• Arch x86, Itanium, Opteron, Alpha, PPC, SPARC,
  Cray X1, NEC SX-6, Blue Gene, etc.
• Network SMP, Myrinet, Quadrics, Infiniband, IBM
  LAPI, MPI, Ethernet, SHMEM, etc.
• Give us a try at http//upc.lbl.gov

11
Summary

• UPC designed to be consistent with C
• Expose memory layout
• Flexible communication with pointers and arrays
• Give users more control to achieve high
  performance
• Berkeley UPC compiler provides an open-source and
  portable implementation
• Hand optimized UPC programs match and often beat
  MPIs performance
• Research goal productive user efficient
  compiler