Unified Parallel C (UPC)

This chapter describes the Cray specific UPC functionality available in CCE, and features of the specification which are implementation defined.

This release supports the UPC Language Specification, Version 1.3. The UPC 1.3 standard is discussed on the UPC specification website, http://code.google.com/p/upc-specification.

This chapter describes the Cray specific UPC functionality available in CCE, and features of the specification which are implementation defined. Also see intro_pgas(7), or refer to the appropriate UPC man page.

Be familiar with UPC and understand the differences between the published UPC Introduction and Language Specification paper and the current UPC specification. If there's no familiarity with UPC, refer to the UPC home page at http://upc.gwu.edu. Under the Publications link, select the Introduction to UPC and Language Specification paper. This paper is slightly outdated but contains valuable information about understanding and using UPC. The UPC home page also contains, under the Documentation link, the UPC Language Specification paper.

UPC allows for explicitly specifying parallel programming through language syntax rather than library functions such as those used in MPI and SHMEM by allowing for the reading and writing of memory of other processes with simple assignment statements. Program synchronization occurs only when explicitly programmed; there is no implied synchronization.

UPC is a dialect of the C language. It is not available in C++.

UPC allows for the maintainence of a view of the program as a collection of threads operating in a common global address space without the burden of the details of how parallelism is implemented on the machine (for example, as shared memory or as a collection of physically distributed memories).

UPC data objects are private to a single thread or shared among all threads of execution. Each thread has a unique memory space that holds its private data objects, and access to a globally-shared memory space that is distributed across the threads. Thus, every part of a shared data object has an affinity to a single thread.

Cray UPC is compatible with MPI. While it may work in some cases, mixing language-based PGAS with SHMEM is not officially supported.

UPC 1.3 supports a parallel I/O model which provides control over file synchronization. However, if use is continued on the regular C I/O routines, supply the controls as needed to remove race conditions. File I/O under UPC is very similar to standard C because one thread opens a file and shares the file handle, and multiple threads may read or write to the same file.

Cray UPC supports GASP instrumentation. GASP instrumentation enables the use of external performance tools, such as the Parallel Performance Wizard (PPW) from the University of Florida. For more information on GASP and PPW, see http://gasp.hcs.ufl.edu and http://ppw.hcs.ufl.edu. To instrument for GASP, refer to the command line option -h [no]gasp[=opt[:opt]].

                 Register     Memory
Initial Value                 0x0000
PE 0 Reads         0x0000     0x0000
PE 1 Reads         0x0000     0x0000
PE 0 Inserts 3     0x3000     0x0000
PE 1 Inserts 7     0x0700     0x0000
PE 0 Writes        0x3000     0x3000
PE 1 Writes        0x0700     0x0700

% cc -hupc hello.c -o hello

% cc -hupc x.o y.o z.o

% ftn x.o y.o z.o

% aprun -n 128 ./hello

% export XT_SYMMETRIC_HEAP_SIZE=512M
% aprun -n 128 ./hello