Zhang Zhaoqing

OpenMP on networks of workstations for software DSMs

This paper describes the implementation of a sizable subset of OpenMP on networks of workstations (NOWs) and the source-to-source OpenMP compiler (AutoPar) is used for the JIAJIA home-based shared virtual memory system (SVM). The paper suggests some simple modifications and extensions to the OpenMP standard for the difference between SVM and SMP (symetric multi processor), at which the OpenMP specification is aimed. The OpenMP translator is based on an automatic parallelization compiler, so it is possible to check the correctness of the semantics of OpenMP programs which is not required in an OpenMP-compliant implementation.AutoPar is measured for five applications including both programs from NAS Parallel Benchmarks and real applications on a cluster of eight Pentium II PCs connected by a 100Mbps switched Ethernet. The evaluation shows that the parallelization by annotating OpenMP directives in simple and the performance of generated JIAJIA code is still acceptable on NOWs.

Advanced compilation techniques used in PORT system

Compilers perform interprocedural parallelization and optimization become an essential component of modern high performance computer systems. A compiler system that is called PORT is presented in this paper. Its important features, such as flow analysis, data dependence analysis and visualization tools are addressed.

Interprocedural constant range propagation and alias analysis by multiple version method

A set of methods for interprocedural analysis is proposed. First, an approach for interprocedural constant propagation is given. Then the concept of constant propagation is extended so as to meet the needs of data dependence analysis. Besides certain constant, constant range can also be propagated. The related propagating rules are introduced, and an idea for computing Return function is given. This approach can solve almost all interprocedural constant propagation problems with non-recursive calls. Second, a multiple-version parallelizing technique is also proposed for alias problem. The work related to this paper has been implemented on a shared-memory parallel computer.

Queue usage and memory-level parallelism sensitive scheduling

In out-of-order (OOO) processors, reorder queue (ROQ) has been widely used to implement precise interruption. The full of ROQ will cause the whole processor stall, while a long latency operation, e.g. a load missed in the caches, will almost definitely cause the ROQ full. In this paper, we present a model for estimating the impact of issuing an instruction on the usage of ROQ and memory level parallelism (MLP), and incorporate these considerations in the cost model of instruction scheduling. Preliminary evaluation results are presented to demonstrate the effectiveness of our approach on reducing the time of ROQ full and improving performance

Multi-view intermediate representation based on algebraic data type

The structure of intermediate representation (IR) is usually a trade-off between several design considerations in practical compiler systems, and may not be appropriate for expressing the algorithms of some particular phases. On the other hand, conventional designs do not distinguish between the logical view and the physical view of IR. The high-level logic of algorithms is always interwoven with the low-level details operating in the IR. So it is difficult to develop or maintain a compiler or to reuse the code of existing compilers. In an attempt to solve these problems, this paper presents a multi-view IR based on an algebraic data type.

Supporting flexible data distribution in software DSMs

Page-based software DSM systems suffer from false sharing caused by the large sharing granularity, and only support one-dimensionBlock orCyclicblock data distribution schemes. Thus applications running on them will suffer from poor data locality and will be able to exploit parallelism only when using a large number of processors. In this paper, a way towards supporting flexible data distribution (FDD) on software DSM system is presented. Small granularity-tunable blocks, the size of which can be set by compiler or programmer, are used to overlap the working data sets distributed among processors. The FDD was implemented on a software DSM system called JIAJIA. Compared withBlock/Cyclic-block distribution schemes used by most DSM systems now, experiments show that the proposed way of flexible data distribution is more effective. The performance of the applications used in the experiments is significantly improved.

Precise dependence test for scalars within nested loops

Exact direction and distance vectors are essential for detecting hierarchical parallelism and examining legality of loop transformation for a multiple level loop nest. Much of this work has been concentrated on array references. Little has been done to address the problems of finding precise dependences between scalar references, except to use extended SSA form with factored use-def links. In this paper, we present a technique for calculating precise direction and distance vectors for scalar references within nested loops without using any forms of SSA. To do this, we use conventional use-def links in combination with joint dominator and joint postdominator relationships, which are extended from dominator and postdominator respectively in standard data flow analysis. The precision of dependence information gathered by our algorithm can not be achieved by traditional analysis of dominator or reaching definitions.