A performance analysis of the Berkeley UPC compiler

Chen, Wei-Yu; Bonachea, Dan; Duell, Jason; Husbands, Parry; Iancu, Costin; Yelick, Katherine

doi:10.1145/782814.782825

Cited by 87 publications

(17 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These PGAS languages distinguish between local and global references and provide support for logical data layouts, such as block-cyclic array distribution. As a result, they implement complicated memory management and a reference to a global object is orders of magnitude slower [18] than a load/store instruction due to complex addressing rules.…”

Section: Unified Parallel Cmentioning

confidence: 99%

Scaling data race detection for partitioned global address space programs

Park

Sen

Iancu

2013

Proceedings of the 27th International ACM Conference on International Conference on Supercomputing

Self Cite

View full text Add to dashboard Cite

Contemporary and future programming languages for HPC promote hybrid parallelism and shared memory abstractions using a global address space. In this programming style, data races occur easily and are notoriously hard to find. Existing state-of-the-art data race detectors exhibit 10 × −100× performance degradation and do not handle hybrid parallelism. In this paper we present the first complete implementation of data race detection at scale for UPC programs. Our implementation tracks local and global memory references in the program and it uses two techniques to reduce the overhead: 1) hierarchical function and instruction level sampling; and 2) exploiting the runtime persistence of aliasing and locality specific to Partitioned Global Address Space applications. The results indicate that both techniques are required in practice: well optimized instruction sampling introduces overheads as high as 6500% (65× slowdown), while each technique in separation is able to reduce it only to 1000% (10× slowdown). When applying the optimizations in conjunction our tool finds all previously known data races in our benchmark programs with at most 50% overhead when running on 2048 cores. Furthermore, while previous results illustrate the benefits of function level sampling, our experiences show that this technique does not work for scientific programs: instruction sampling or a hybrid approach is required.

show abstract

Section: Unified Parallel Cmentioning

confidence: 99%

Scaling data race detection for partitioned global address space programs

Park

Sen

Iancu

2013

Proceedings of the 27th International ACM Conference on International Conference on Supercomputing

Self Cite

View full text Add to dashboard Cite

show abstract

“…In this sub-section we compare the programmability for this dynamic process reconfiguration between several representative data communication models: a message passing model, a local-view-based PGAS model [28], [29], a global-view-based PGAS model [16], [23] and a DSM (Distributed Shared Memory) model [5], [18].…”

Section: Data Communication Modelsmentioning

confidence: 99%

“…Second, in the global-view-based PGAS model such as UPC [16], Global Arrays [23], X10 [4] and Chapel [6], the programmer can describe data communications as read/write operations from/to a global address space. This global-view-based PGAS model is very suitable to program the dynamic process reconfiguration.…”

Section: Data Communication Modelsmentioning

confidence: 99%

“…However, DMI adopts the globalview-based PGAS model instead of the DSM model since the PGAS model provides the programmer with more explicit ways for controlling data distribution to optimize his program straightforwardly. Finally, note that regardless of the reconfiguration, the global-view-based PGAS model delivers good programmability based on read/write operations similar to multi-thread programming on a physically shared memory environment [16], [23].…”

Section: Data Communication Modelsmentioning

confidence: 99%

“…Most existing PGAS frameworks such as Co-Array Fortran [28], UPC [16], Global Arrays [23], X10 [4] and Chapel [6] support only get/put operations and do not support cache and dynamic change of data c 2012 Information Processing Society of Japan distribution. In contrast, most existing DSM frameworks such as TreadMarks [5], JIAJIA [13] and DSM-Threads [22] support cache but a programmer cannot freely mix the invalidate protocol and the update protocol.…”

Section: Related Workmentioning

confidence: 99%

See 2 more Smart Citations

Parallel Computational Reconfiguration Based on a PGAS Model

Hara

Taura

2012

Journal of Information Processing

View full text Add to dashboard Cite

Abstract:In order to improve the resource utilization of clusters and supercomputers and thus deliver application results to users faster, it is essential for a job scheduler to expand and shrink parallel computations flexibly. In order to enable the flexible job scheduling, the parallel computations have to be reconfigurable. With this motivation, this paper proposes, implements and evaluates DMI, a global-view-based PGAS framework that enables easy programming of reconfigurable and high-performance parallel iterative computations. DMI provides programming interfaces with which a programmer can program the reconfiguration easily with a global-view. Our performance evaluations showed that DMI can efficiently adapt the parallelism of long-running parallel iterative computations, such as a real-world finite element method and large-scale iterative graph search, to the dynamic increase and decrease of available resources through the reconfiguration.

show abstract

UPCBLAS: a library for parallel matrix computations in Unified Parallel C

González-Domínguez

Martín

Taboada

et al. 2012

Concurrency and Computation

View full text Add to dashboard Cite

The popularity of Partitioned Global Address Space (PGAS) languages has increased during the last years thanks to their high programmability and performance through an efficient exploitation of data locality, especially on hierarchical architectures such as multicore clusters. This paper describes UP-CBLAS, a parallel numerical library for dense matrix computations using the PGAS Unified Parallel C (UPC) language. The routines developed in UPCBLAS are built on top of sequential BLAS functions and exploit the particularities of the PGAS paradigm, taking into account data locality in order to achieve a good performance. Furthermore, the routines implement other optimization techniques, several of them by automatically taking into account the hardware characteristics of the underlying systems on which they are executed. The library has been experimentally evaluated on a multicore supercomputer and compared to a message-passing based parallel numerical library, demonstrating good scalability and efficiency.

show abstract

A performance analysis of the Berkeley UPC compiler

Abstract: Unified Parallel C (UPC) is a parallel language that uses a Sin

Cited by 87 publications

References 13 publications

Scaling data race detection for partitioned global address space programs

Scaling data race detection for partitioned global address space programs

Parallel Computational Reconfiguration Based on a PGAS Model

UPCBLAS: a library for parallel matrix computations in Unified Parallel C

Contact Info

Product

Resources

About