In this paper, we describe efficient, scalable, and deadlock-free asynchronous communication strategies suitable for unstructured computations on iPSC/860. Using these deadlock-free strategies, which incur small overhead, we have optimized the communication in parallel sparse Cholesky factorization. We present experimental results to show that such optimizations can lead to more than 50% saving in communication costs.

The authors propose a Cholesky factorization scheme based on a static task and communication schedule generated by symbolic preprocessing of the intercolumn dependencies. This information is used to reduce the overheads of maintaining data structures and the communication costs incurred in message passing during the numerical factorization step. The authors introduce three primitives which encapsulate these optimizations and alleviate the user's programming effort. Performance results obtained on the iPSC/860 show that an implementation using these primitives results in 30% to 40% savings over one that does not use any static communication structure information.>

The problem of cache thrashing occurs in shared memory multiprocessors with local processor caches and a write-invalidate cache coherency protocol.Cache conflict, one of the causes of thrashing, arises when several processors try to write into different addresses of the same cache line.It is shown that parallel execution of loop iterations on a machine with a long cache line can cause significantly more thrashing than identical execution on a machine with a cache line size of one or two words.Transformations that can be applied to :parallel loops to reduce cache conflicts and a method fcm trading off this reduction against the possible increase in synchronization and scheduling overheads are described.The applicability of this technique is shown by an empirical analysis of FORTRAN loops in the Perfect Club benchmark suite.

A block based, automatic partitioning and scheduling methodology is presented for sparse matrix factorization on distributed memory systems. Using experimental results, this technique is analyzed for communication and load imbalance overhead. To study the performance effects, these overheads were compared with those obtained from a straightforward 'wrap mapped' column assignment scheme. All experimental results were obtained using test sparse matrices from the Harwell-Boeing data set. The results show that there is a communication and load balance tradeoff. The block based method results in lower communication cost whereas the wrap mapped scheme gives better load balance.

Clustering is a mapping of the nodes of a task graph onto labeled clusters. We present a unified framework for clustering of directed acyclic graphs (DAGs). Several clustering algorithms from the literature are compared using this framework. For coarse grain DAGs two interesting properties are presented. For every nonlinear clustering there exists a linear clustering whose parallel time is less than the nonlinear one. Furthermore, the parallel time of any linear clustering is within a factor of two of the optimal. Two clustering algorithms are presented with near linear time complexity for coarse grain DAGs. The conclusion is that linear clustering is an efficient and accurate operation.