No abstract

Combinatorial problems such as those from graph theory pose serious challenges for parallel machines due to non-contiguous, concurrent

No abstract

This paper introduces Eldorado, a third generation multithreaded architecture. Previous Cray multithreaded systems were plagued by unreliable hardware and high costs. Eldorado corrects these problems by using many parts built for other commercial systems. Its compute processor is a 500 MHZ multithreaded processor architecturally similar to the MTA-2 processor; but its interconnection network, I/O subsystem, and service processors are borrowed from other Cray systems. Eldorado retains the programming model, operating system, and tools of the MTA-2. It has the same capability as the MTA-2 to tolerate latencies and achieve high performance on programs that run poorly on SMP clusters. We present several programming examples to illustrate performance and scalability in the presence of high memory and synchronization latencies.

Abstract-It is our view that the state of the art in constructing a large collection of graph algorithms in terms of linear algebraic operations is mature enough to support the emergence of a standard set of primitive building blocks. This paper is a position paper defining the problem and announcing our intention to launch an open effort to define this standard.Keywords-component; Graphs; Algorithms; Linear Algebra; Software Standards I.PROBLEM STATEMENT Data analytics and the closely related field of "big data" have emerged as a leading research topics in both applied and theoretical computer science. While it has been shown that many problems can be addressed with a "map-reduce" style framework, as we move to the next level of sophistication in data analytics applications, graph algorithms that demand more than "map-reduce" will play an increasingly vital role. There are many ways to organize a collection of graph algorithms into a high level library to support data analytics. It is probably premature to standardize these graph APIs. The low level building blocks of graph algorithms, however, are well understood and we believe a suitable target for standardization. In particular, the representation of graphs as sparse matrices allows many graph algorithms to be represented in terms of a modest set of linear algebra operations [1,2,5].Our concern, however, is that as new researchers enter this expanding field of research, the linear algebraic foundation of this class of graph algorithms will fragment. Diversity at the level of the primitive building blocks of graph algorithms will not help advance the field of graph algorithms. It will hinder progress as groups create different overlapping variants of what should be common low level building blocks. Furthermore, diverse sets of primitives will complicate the ability of the vendor community to support this research with math tuned to the needs of these algorithms.It is our view that the state of the art in constructing a large collection of graph algorithms in terms of linear algebraic operations is mature enough to support the emergence of a standard set of primitive building blocks. We believe it is critical that we move quickly so as new research groups enter this field we can prevent needless and ultimately damaging diversity at the level of the basic primitives supporting this research; thereby freeing up researchers to innovate and diversify at the level of higher level algorithms and graph analytics applications. II. THE STATE-OF-THE-ARTThe standardization of sparse linear algebra historically begins with the NIST Sparse Basic Linear Algebra Subprograms (BLAS) [3] and consists of Sparse Vector (Level 1), Matrix Vector (Level 2), and Matrix Matrix (Level 3) operations. These BLAS were designed for solving the kinds of sparse linear algebra operations that arise in finite element simulation techniques that are widely used in engineering. In particular, the operations are limited to traditional multiplication and addition operations and, in the case of mat...