Edgar Solomonik

Improving communication performance in dense linear algebra via topology aware collectives

Recent results have shown that topology aware mapping reduces network contention in communication-intensive kernels on massively parallel machines. We demonstrate that on mesh interconnects, topology aware mapping also allows for the utilization of highly-efficient topology aware collectives. We map novel 2.5D dense linear algebra algorithms to exploit rectangular collectives on cuboid partitions allocated by a Blue Gene/P supercomputer. Our mappings allow the algorithms to exploit optimized line multicasts and reductions. Commonly used 2D algorithms cannot be mapped in this fashion. On 16,384 nodes (65,536 cores) of Blue Gene/P, 2.5D algorithms that exploit rectangular collectives are sig- nificantly faster than 2D matrix multiplication (MM) and LU factorization, up to 8.7x and 2.1x, respectively. These speed-ups are due to communication reduction (up to 95.6% for 2.5D MM with respect to 2D MM). We also derive LogP- based novel performance models for rectangular broadcasts and reductions. Using those, we model the performance of matrix multiplication and LU factorization on a hypothetical exascale architecture.

Highly scalable parallel sorting

Sorting is a commonly used process with a wide breadth of applications in the high performance computing field. Early research in parallel processing has provided us with comprehensive analysis and theory for parallel sorting algorithms. However, modern supercomputers have advanced rapidly in size and changed significantly in architecture, forcing new adaptations to these algorithms. To fully utilize the potential of highly parallel machines, tens of thousands of processors are used. Efficiently scaling parallel sorting on machines of this magnitude is inhibited by the communication-intensive problem of migrating large amounts of data between processors. The challenge is to design a highly scalable sorting algorithm that uses minimal communication, maximizes overlap between computation and communication, and uses memory efficiently. This paper presents a scalable extension of the Histogram Sorting method, making fundamental modifications to the original algorithm in order to minimize message contention and exploit overlap. We implement Histogram Sort, Sample Sort, and Radix Sort in Charm++ and compare their performance. The choice of algorithm as well as the importance of the optimizations is validated by performance tests on two predominant modern supercomputer architectures: XT4 at ORNL (Jaguar) and Blue Gene/P at ANL (Intrepid).

A massively parallel tensor contraction framework for coupled-cluster computations

Precise calculation of molecular electronic wavefunctions by methods such as coupled-cluster requires the computation of tensor contractions, the cost of which has polynomial computational scaling with respect to the system and basis set sizes. Each contraction may be executed via matrix multiplication on a properly ordered and structured tensor. However, data transpositions are often needed to reorder the tensors for each contraction. Writing and optimizing distributed-memory kernels for each transposition and contraction is tedious since the number of contractions scales combinatorially with the number of tensor indices. We present a distributed-memory numerical library (Cyclops Tensor Framework (CTF)) that automatically manages tensor blocking and redistribution to perform any user-specified contractions. CTF serves as the distributed-memory contraction engine in Aquarius, a new program designed for high-accuracy and massively-parallel quantum chemical computations. Aquarius implements a range of coupled-cluster and related methods such as CCSD and CCSDT by writing the equations on top of a C++ templated domain-specific language. This DSL calls CTF directly to manage the data and perform the contractions. Our CCSD and CCSDT implementations achieve high parallel scalability on the BlueGene/Q and Cray XC30 supercomputer architectures showing that accurate electronic structure calculations can be effectively carried out on top of general distributed-memory tensor primitives. We introduce Cyclops Tensor Framework (CTF), a distributed-memory library for tensor contractions.CTF is able to perform tensor decomposition, redistribution, and contraction at runtime.CTF enables the expression of massively-parallel coupled-cluster methods via a concise tensor contraction interface.The quantum chemistry software suite Aquarius employs CTF to execute two coupled-cluster methods: CCSD and CCSDT.The Aquarius CCSD and CCSDT codes scale well on BlueGene/Q and Cray XC30, comparing favorably to NWChem.

Understanding Application Performance via Micro-benchmarks on Three Large Supercomputers: Intrepid, Ranger and Jaguar

The emergence of new parallel architectures presents new challenges for application developers. Supercomputers vary in processor speed, network topology, interconnect communication characteristics and memory subsystems. This paper presents a performance comparison of three of the fastest machines in the world: IBM’s Blue Gene/P installation at ANL (Intrepid), the SUN-Infiniband cluster at TACC (Ranger) and Cray’s XT4 installation at ORNL (Jaguar). Comparisons are based on three applications selected by NSF for the Track 1 proposal to benchmark the Blue Waters system: NAMD, MILC and a turbulence code, DNS. We present a comprehensive overview of the architectural details of each of these machines and a comparison of their basic performance parameters. Application performance is presented for multiple problem sizes and the relative performance on the selected machines is explained through micro-benchmarking results. We hope that insights from this work will be useful to managers making buying decisions for supercomputers and application users trying to decide on a machine to run on. Based on the performance analysis techniques used in the paper, we also suggest a step-by-step procedure for estimating the suitability of a given architecture for a highly parallel application.

Communication avoiding and overlapping for numerical linear algebra

To efficiently scale dense linear algebra problems to future exascale systems, communication cost must be avoided or overlapped. Communication-avoiding 2.5D algorithms improve scalability by reducing inter-processor data transfer volume at the cost of extra memory usage. Communication overlap attempts to hide messaging latency by pipelining messages and overlapping with computational work. We study the interaction and compatibility of these two techniques for two matrix multiplication algorithms (Cannon and SUMMA), triangular solve, and Cholesky factorization. For each algorithm, we construct a detailed performance model that considers both critical path dependencies and idle time. We give novel implementations of 2.5D algorithms with overlap for each of these problems. Our software employs UPC, a partitioned global address space (PGAS) language that provides fast one-sided communication. We show communication avoidance and overlap provide a cumulative benefit as core counts scale, including results using over 24K cores of a Cray XE6 system.

Tradeoffs between synchronization, communication, and computation in parallel linear algebra computations

This paper derives tradeoffs between three basic costs of a parallel algorithm: synchronization, data movement, and computational cost. These tradeoffs are lower bounds on the execution time of the algorithm which are independent of the number of processors, but dependent on the problem size. Therefore, they provide lower bounds on the parallel execution time of any algorithm computed by a system composed of any number of homogeneous components, each with associated computational, communication, and synchronization payloads. We employ a theoretical model counts the amount of work and data movement as a maximum of any execution path during the parallel computation. By considering this metric, rather than the total communication volume over the whole machine, we obtain new insights into the characteristics of parallel schedules for algorithms with non-trivial dependency structures. We also present reductions from BSP and LogP algorithms to our execution model, extending our lower bounds to these two models of parallel computation. We first develop our results for general dependency graphs and hypergraphs based on their expansion properties, then we apply the theorem to a number of specific algorithms in numerical linear algebra, namely triangular substitution, Gaussian elimination, and Krylov subspace methods. Our lower bound for LU factorization demonstrates the optimality of Tiskins LU algorithm answering an open question posed in his paper, as well as of the 2.5D LU algorithm which has analogous costs. We treat the computations in a general manner by noting that the computations share a similar dependency hypergraph structure and analyzing the communication requirements of lattice hypergraph structures.

Minimizing Communication in All-Pairs Shortest Paths

We consider distributed memory algorithms for the all-pairs shortest paths (APSP) problem. Scaling the APSP problem to high concurrencies requires both minimizing inter-processor communication as well as maximizing temporal data locality. The 2.5D APSP algorithm, which is based on the divide-and-conquer paradigm, satisfies both of these requirements: it can utilize any extra available memory to perform asymptotically less communication, and it is rich in semiring matrix multiplications, which have high temporal locality. We start by introducing a block-cyclic 2D (minimal memory) APSP algorithm. With a careful choice of block-size, this algorithm achieves known communication lower-bounds for latency and bandwidth. We extend this 2D block-cyclic algorithm to a 2.5D algorithm, which can use c extra copies of data to reduce the bandwidth cost by a factor of c1/2, compared to its 2D counterpart. However, the 2.5D algorithm increases the latency cost by c1/2. We provide a tighter lower bound on latency, which dictates that the latency overhead is necessary to reduce bandwidth along the critical path of execution. Our implementation achieves impressive performance and scaling to 24,576 cores of a Cray XE6 supercomputer by utilizing well-tuned intra-node kernels within the distributed memory algorithm.

A Communication-Optimal N-Body Algorithm for Direct Interactions

We consider the problem of communication avoidance in computing interactions between a set of particles in scenarios with and without a cutoff radius for interaction. Our strategy, which we show to be optimal in communication, divides the work in the iteration space rather than simply dividing the particles over processors, so more than one processor may be responsible for computing updates to a single particle. Similar to a force decomposition in molecular dynamics, this approach requires up to √p times more memory than a particle decomposition, but reduces communication costs by factors up to √p and is often faster in practice than a particle decomposition [1]. We examine a generalized force decomposition algorithm that tolerates the memory limited case, i.e. when memory can only hold c copies of the particles for c = 1, 2, ..., √p. When c = 1, the algorithm degenerates into a particle decomposition; similarly when c = √p, the algorithm uses a force decomposition. We present a proof that the algorithm is communication-optimal and reduces critical path latency and bandwidth costs by factors of c2 and c, respectively. Performance results from experiments on up to 24K cores of Cray XE-6 and 32K cores of IBM BlueGene/P machines indicate that the algorithm reduces communication in practice. In some cases, it even outperforms the original force decomposition approach because the right choice of c strikes a balance between the costs of collective and point-to-point communication. Finally, we extend the analysis to include a cutoff radius for direct evaluation of force interactions. We show that with a cutoff, communication optimality still holds. We sketch a generalized algorithm for multi-dimensional space and assess its performance for 1D and 2D simulations on the same systems.

Reconstructing Householder Vectors from Tall-Skinny QR

The Tall-Skinny QR (TSQR) algorithm is more communication efficient than the standard Householder algorithm for QR decomposition of matrices with many more rows than columns. However, TSQR produces a different representation of the orthogonal factor and therefore requires more software development to support the new representation. Further, implicitly applying the orthogonal factor to the trailing matrix in the context of factoring a square matrix is more complicated and costly than with the Householder representation. We show how to perform TSQR and then reconstruct the Householder vector representation with the same asymptotic communication efficiency and little extra computational cost. We demonstrate the high performance and numerical stability of this algorithm both theoretically and empirically. The new Householder reconstruction algorithm allows us to design more efficient parallel QR algorithms, with significantly lower latency cost compared to Householder QR and lower bandwidth and latency costs compared with Communication-Avoiding QR (CAQR) algorithm. As a result, our final parallel QR algorithm outperforms ScaLAPACK and Elemental implementations of Householder QR and our implementation of CAQR on the Hopper Cray XE6 NERSC system. We also provide algorithmic improvements to the ScaLAPACK and CAQR algorithms.

To Push or To Pull: On Reducing Communication and Synchronization in Graph Computations

We reduce the cost of communication and synchronization in graph processing by analyzing the fastest way to process graphs: pushing the updates to a shared state or pulling the updates to a private state. We investigate the applicability of this push-pull dichotomy to various algorithms and its impact on complexity, performance, and the amount of used locks, atomics, and reads/writes. We consider 11 graph algorithms, 3 programming models, 2 graph abstractions, and various families of graphs. The conducted analysis illustrates surprising differences between push and pull variants of different algorithms in performance, speed of convergence, and code complexity; the insights are backed up by performance data from hardware counters. We use these findings to illustrate which variant is faster for each algorithm and to develop generic strategies that enable even higher speedups. Our insights can be used to accelerate graph processing engines or libraries on both massively-parallel shared-memory machines as well as distributed-memory systems.