arXiv subjects
Publications and source records attributed to James Demmel.
Interprocessor communication often dominates the runtime of large matrix computations. We present a parallel algorithm for computing QR decompositions whose bandwidth cost (communication volume) can be decreased at the cost of increasing its latency cost (number of messages). By varying a parameter to navigate the bandwidth/latency tradeoff, we can tune this algorithm for machines with different communication costs.
We propose two new methods to address the weak scaling problems of KRR: the Balanced KRR (BKRR) and K-means KRR (KKRR). These methods consider alternative ways to partition the input dataset into p different parts, generating p different models, and then selecting the best model among them. Compared to a conventional implementation, KKRR2 (optimized version of KKRR) improves the weak scaling efficiency from 0.32% to 38% and achieves a 591times speedup for getting the same accuracy by using the same data and the same hardware (1536 processors). BKRR2 (optimized version of BKRR) achieves a higher accuracy than the current fastest method using less training time for a variety of datasets. For the applications requiring only approximate solutions, BKRR2 improves the weak scaling efficiency to 92% and achieves 3505 times speedup (theoretical speedup: 4096 times).
Efficiently executing convolutional neural nets (CNNs) is important in many machine-learning tasks. Since the cost of moving a word of data, either between levels of a memory hierarchy or between processors over a network, is much higher than the cost of an arithmetic operation, minimizing data movement is critical to performance optimization. In this paper, we present both new lower bounds on data movement needed for CNNs, and optimal sequential algorithms that attain these lower bounds. In most common cases, our optimal algorithms can attain significantly more data reuse than matrix multiplication.
Finishing 90-epoch ImageNet-1k training with ResNet-50 on a NVIDIA M40 GPU takes 14 days. This training requires 10^18 single precision operations in total. On the other hand, the world's current fastest supercomputer can finish 2 * 10^17 single precision operations per second (Dongarra et al 2017, https://www.top500.org/lists/2017/06/). If we can make full use of the supercomputer for DNN training, we should be able to finish the 90-epoch ResNet-50 training in one minute. However, the current bottleneck for fast DNN training is in the algorithm level. Specifically, the current batch size (e.g. 512) is too small to make efficient use of many processors. For large-scale DNN training, we focus on using large-batch data-parallelism synchronous SGD without losing accuracy in the fixed epochs. The LARS algorithm (You, Gitman, Ginsburg, 2017, arXiv:1708.03888) enables us to scale the batch size to extremely large case (e.g. 32K). We finish the 100-epoch ImageNet training with AlexNet in 11 minutes on 1024 CPUs. About three times faster than Facebook's result (Goyal et al 2017, arXiv:1706.02677), we finish the 90-epoch ImageNet training with ResNet-50 in 20 minutes on 2048 KNLs without losing accuracy. State-of-the-art ImageNet training speed with ResNet-50 is 74.9% top-1 test accuracy in 15 minutes. We got 74.9% top-1 test accuracy in 64 epochs, which only needs 14 minutes. Furthermore, when we increase the batch size to above 16K, our accuracy is much higher than Facebook's on corresponding batch sizes. Our source code is available upon request.
Parallel computing has played an important role in speeding up convex optimization methods for big data analytics and large-scale machine learning (ML). However, the scalability of these optimization methods is inhibited by the cost of communicating and synchronizing processors in a parallel setting. Iterative ML methods are particularly sensitive to communication cost since they often require communication every iteration. In this work, we extend well-known techniques from Communication-Avoiding Krylov subspace methods to first-order, block coordinate descent methods for Support Vector Machines and Proximal Least-Squares problems. Our Synchronization-Avoiding (SA) variants reduce the latency cost by a tunable factor of $s$ at the expense of a factor of $s$ increase in flops and bandwidth costs. We show that the SA-variants are numerically stable and can attain large speedups of up to $5.1\times$ on a Cray XC30 supercomputer.
The fast iterative soft thresholding algorithm (FISTA) is used to solve convex regularized optimization problems in machine learning. Distributed implementations of the algorithm have become popular since they enable the analysis of large datasets. However, existing formulations of FISTA communicate data at every iteration which reduces its performance on modern distributed architectures. The communication costs of FISTA, including bandwidth and latency costs, is closely tied to the mathematical formulation of the algorithm. This work reformulates FISTA to communicate data at every k iterations and reduce data communication when operating on large data sets. We formulate the algorithm for two different optimization methods on the Lasso problem and show that the latency cost is reduced by a factor of k while bandwidth and floating-point operation costs remain the same. The convergence rates and stability properties of the reformulated algorithms are similar to the standard formulations. The performance of communication-avoiding FISTA and Proximal Newton methods is evaluated on 1 to 1024 nodes for multiple benchmarks and demonstrate average speedups of 3-10x with scaling properties that outperform the classical algorithms.
The speed of deep neural networks training has become a big bottleneck of deep learning research and development. For example, training GoogleNet by ImageNet dataset on one Nvidia K20 GPU needs 21 days. To speed up the training process, the current deep learning systems heavily rely on the hardware accelerators. However, these accelerators have limited on-chip memory compared with CPUs. To handle large datasets, they need to fetch data from either CPU memory or remote processors. We use both self-hosted Intel Knights Landing (KNL) clusters and multi-GPU clusters as our target platforms. From an algorithm aspect, current distributed machine learning systems are mainly designed for cloud systems. These methods are asynchronous because of the slow network and high fault-tolerance requirement on cloud systems. We focus on Elastic Averaging SGD (EASGD) to design algorithms for HPC clusters. Original EASGD used round-robin method for communication and updating. The communication is ordered by the machine rank ID, which is inefficient on HPC clusters. First, we redesign four efficient algorithms for HPC systems to improve EASGD's poor scaling on clusters. Async EASGD, Async MEASGD, and Hogwild EASGD are faster \textcolor{black}{than} their existing counterparts (Async SGD, Async MSGD, and Hogwild SGD, resp.) in all the comparisons. Finally, we design Sync EASGD, which ties for the best performance among all the methods while being deterministic. In addition to the algorithmic improvements, we use some system-algorithm codesign techniques to scale up the algorithms. By reducing the percentage of communication from 87% to 14%, our Sync EASGD achieves 5.3x speedup over original EASGD on the same platform. We get 91.5% weak scaling efficiency on 4253 KNL cores, which is higher than the state-of-the-art implementation.
Primal and dual block coordinate descent methods are iterative methods for solving regularized and unregularized optimization problems. Distributed-memory parallel implementations of these methods have become popular in analyzing large machine learning datasets. However, existing implementations communicate at every iteration which, on modern data center and supercomputing architectures, often dominates the cost of floating-point computation. Recent results on communication-avoiding Krylov subspace methods suggest that large speedups are possible by re-organizing iterative algorithms to avoid communication. We show how applying similar algorithmic transformations can lead to primal and dual block coordinate descent methods that only communicate every $s$ iterations--where $s$ is a tuning parameter--instead of every iteration for the \textit{regularized least-squares problem}. We show that the communication-avoiding variants reduce the number of synchronizations by a factor of $s$ on distributed-memory parallel machines without altering the convergence rate and attains strong scaling speedups of up to $6.1\times$ on a Cray XC30 supercomputer.
This work studies the application of the discrete Holder-Brascamp-Lieb (HBL) inequalities to the design of communication optimal algorithms. In particular, it describes optimal tiling (blocking) strategies for nested loops that lack data dependencies and exhibit linear memory access patterns. We attain known lower bounds for communication costs by unraveling the relationship between the HBL linear program, its dual, and tile selection. The methods used are constructive and algorithmic. The case when all arrays have one index is explored in depth, as a useful example in which a particularly efficient tiling can be determined.
Sparse matrix-matrix multiplication (or SpGEMM) is a key primitive for many high-performance graph algorithms as well as for some linear solvers, such as algebraic multigrid. The scaling of existing parallel implementations of SpGEMM is heavily bound by communication. Even though 3D (or 2.5D) algorithms have been proposed and theoretically analyzed in the flat MPI model on Erdos-Renyi matrices, those algorithms had not been implemented in practice and their complexities had not been analyzed for the general case. In this work, we present the first ever implementation of the 3D SpGEMM formulation that also exploits multiple (intra-node and inter-node) levels of parallelism, achieving significant speedups over the state-of-the-art publicly available codes at all levels of concurrencies. We extensively evaluate our implementation and identify bottlenecks that should be subject to further research.
We explore the trade-offs of performing linear algebra using Apache Spark, compared to traditional C and MPI implementations on HPC platforms. Spark is designed for data analytics on cluster computing platforms with access to local disks and is optimized for data-parallel tasks. We examine three widely-used and important matrix factorizations: NMF (for physical plausability), PCA (for its ubiquity) and CX (for data interpretability). We apply these methods to TB-sized problems in particle physics, climate modeling and bioimaging. The data matrices are tall-and-skinny which enable the algorithms to map conveniently into Spark's data-parallel model. We perform scaling experiments on up to 1600 Cray XC40 nodes, describe the sources of slowdowns, and provide tuning guidance to obtain high performance.
Many large-scale scientific computations require eigenvalue solvers in a scaling regime where efficiency is limited by data movement. We introduce a parallel algorithm for computing the eigenvalues of a dense symmetric matrix, which performs asymptotically less communication than previously known approaches. We provide analysis in the Bulk Synchronous Parallel (BSP) model with additional consideration for communication between a local memory and cache. Given sufficient memory to store $c$ copies of the symmetric matrix, our algorithm requires $Θ(\sqrt{c})$ less interprocessor communication than previously known algorithms, for any $c\leq p^{1/3}$ when using $p$ processors. The algorithm first reduces the dense symmetric matrix to a banded matrix with the same eigenvalues. Subsequently, the algorithm employs successive reduction to $O(\log p)$ thinner banded matrices. We employ two new parallel algorithms that achieve lower communication costs for the full-to-band and band-to-band reductions. Both of these algorithms leverage a novel QR factorization algorithm for rectangular matrices.
Hölder-Brascamp-Lieb inequalities provide upper bounds for a class of multilinear expressions, in terms of $L^p$ norms of the functions involved. They have been extensively studied for functions defined on Euclidean spaces. Bennett-Carbery-Christ-Tao have initiated the study of these inequalities for discrete Abelian groups and, in terms of suitable data, have characterized the set of all tuples of exponents for which such an inequality holds for specified data, as the convex polyhedron defined by a particular finite set of affine inequalities. In this paper we advance the theory of such inequalities for torsion-free discrete Abelian groups in three respects. The optimal constant in any such inequality is shown to equal $1$ whenever it is finite. An algorithm that computes the admissible polyhedron of exponents is developed. It is shown that nonetheless, existence of an algorithm that computes the full list of inequalities in the Bennett-Carbery-Christ-Tao description of the admissible polyhedron for all data, is equivalent to an affirmative solution of Hilbert's Tenth Problem over the rationals. That problem remains open. Applications to computer science will be explored in a forthcoming companion paper.
The movement of data (communication) between levels of a memory hierarchy, or between parallel processors on a network, can greatly dominate the cost of computation, so algorithms that minimize communication are of interest. Motivated by this, attainable lower bounds for the amount of communication required by algorithms were established by several groups for a variety of algorithms, including matrix computations. Prior work of Ballard-Demmel-Holtz-Schwartz relied on a geometric inequality of Loomis and Whitney for this purpose. In this paper the general theory of discrete multilinear Holder-Brascamp-Lieb (HBL) inequalities is used to establish communication lower bounds for a much wider class of algorithms. In some cases, algorithms are presented which attain these lower bounds. Several contributions are made to the theory of HBL inequalities proper. The optimal constant in such an inequality for torsion-free Abelian groups is shown to equal one whenever it is finite. Bennett-Carbery-Christ-Tao had characterized the tuples of exponents for which such an inequality is valid as the convex polyhedron defined by a certain finite list of inequalities. The problem of constructing an algorithm to decide whether a given inequality is on this list, is shown to be equivalent to Hilbert's Tenth Problem over the rationals, which remains open. Nonetheless, an algorithm which computes the polyhedron itself is constructed.
The QR factorization and the SVD are two fundamental matrix decompositions with applications throughout scientific computing and data analysis. For matrices with many more rows than columns, so-called "tall-and-skinny matrices," there is a numerically stable, efficient, communication-avoiding algorithm for computing the QR factorization. It has been used in traditional high performance computing and grid computing environments. For MapReduce environments, existing methods to compute the QR decomposition use a numerically unstable approach that relies on indirectly computing the Q factor. In the best case, these methods require only two passes over the data. In this paper, we describe how to compute a stable tall-and-skinny QR factorization on a MapReduce architecture in only slightly more than 2 passes over the data. We can compute the SVD with only a small change and no difference in performance. We present a performance comparison between our new direct TSQR method, a standard unstable implementation for MapReduce (Cholesky QR), and the classic stable algorithm implemented for MapReduce (Householder QR). We find that our new stable method has a large performance advantage over the Householder QR method. This holds both in a theoretical performance model as well as in an actual implementation.
Graph expansion analysis of computational DAGs is useful for obtaining communication cost lower bounds where previous methods, such as geometric embedding, are not applicable. This has recently been demonstrated for Strassen's and Strassen-like fast square matrix multiplication algorithms. Here we extend the expansion analysis approach to fast algorithms for rectangular matrix multiplication, obtaining a new class of communication cost lower bounds. These apply, for example to the algorithms of Bini et al. (1979) and the algorithms of Hopcroft and Kerr (1971). Some of our bounds are proved to be optimal.
Parallel matrix multiplication is one of the most studied fundamental problems in distributed and high performance computing. We obtain a new parallel algorithm that is based on Strassen's fast matrix multiplication and minimizes communication. The algorithm outperforms all known parallel matrix multiplication algorithms, classical and Strassen-based, both asymptotically and in practice. A critical bottleneck in parallelizing Strassen's algorithm is the communication between the processors. Ballard, Demmel, Holtz, and Schwartz (SPAA'11) prove lower bounds on these communication costs, using expansion properties of the underlying computation graph. Our algorithm matches these lower bounds, and so is communication-optimal. It exhibits perfect strong scaling within the maximum possible range. Benchmarking our implementation on a Cray XT4, we obtain speedups over classical and Strassen-based algorithms ranging from 24% to 184% for a fixed matrix dimension n=94080, where the number of nodes ranges from 49 to 7203. Our parallelization approach generalizes to other fast matrix multiplication algorithms.
A parallel algorithm has perfect strong scaling if its running time on P processors is linear in 1/P, including all communication costs. Distributed-memory parallel algorithms for matrix multiplication with perfect strong scaling have only recently been found. One is based on classical matrix multiplication (Solomonik and Demmel, 2011), and one is based on Strassen's fast matrix multiplication (Ballard, Demmel, Holtz, Lipshitz, and Schwartz, 2012). Both algorithms scale perfectly, but only up to some number of processors where the inter-processor communication no longer scales. We obtain a memory-independent communication cost lower bound on classical and Strassen-based distributed-memory matrix multiplication algorithms. These bounds imply that no classical or Strassen-based parallel matrix multiplication algorithm can strongly scale perfectly beyond the ranges already attained by the two parallel algorithms mentioned above. The memory-independent bounds and the strong scaling bounds generalize to other algorithms.