@inproceedings{6558, abstract = {This paper studies the problem of distributed stochastic optimization in an adversarial setting where, out of m machines which allegedly compute stochastic gradients every iteration, an α-fraction are Byzantine, and may behave adversarially. Our main result is a variant of stochastic gradient descent (SGD) which finds ε-approximate minimizers of convex functions in T=O~(1/ε²m+α²/ε²) iterations. In contrast, traditional mini-batch SGD needs T=O(1/ε²m) iterations, but cannot tolerate Byzantine failures. Further, we provide a lower bound showing that, up to logarithmic factors, our algorithm is information-theoretically optimal both in terms of sample complexity and time complexity.}, author = {Alistarh, Dan-Adrian and Allen-Zhu, Zeyuan and Li, Jerry}, booktitle = {Advances in Neural Information Processing Systems}, location = {Montreal, Canada}, pages = {4613--4623}, publisher = {Neural Information Processing Systems Foundation}, title = {{Byzantine stochastic gradient descent}}, volume = {2018}, year = {2018}, } @inproceedings{7116, abstract = {Training deep learning models has received tremendous research interest recently. In particular, there has been intensive research on reducing the communication cost of training when using multiple computational devices, through reducing the precision of the underlying data representation. Naturally, such methods induce system trade-offs—lowering communication precision could de-crease communication overheads and improve scalability; but, on the other hand, it can also reduce the accuracy of training. In this paper, we study this trade-off space, and ask:Can low-precision communication consistently improve the end-to-end performance of training modern neural networks, with no accuracy loss?From the performance point of view, the answer to this question may appear deceptively easy: compressing communication through low precision should help when the ratio between communication and computation is high. However, this answer is less straightforward when we try to generalize this principle across various neural network architectures (e.g., AlexNet vs. ResNet),number of GPUs (e.g., 2 vs. 8 GPUs), machine configurations(e.g., EC2 instances vs. NVIDIA DGX-1), communication primitives (e.g., MPI vs. NCCL), and even different GPU architectures(e.g., Kepler vs. Pascal). Currently, it is not clear how a realistic realization of all these factors maps to the speed up provided by low-precision communication. In this paper, we conduct an empirical study to answer this question and report the insights.}, author = {Grubic, Demjan and Tam, Leo and Alistarh, Dan-Adrian and Zhang, Ce}, booktitle = {Proceedings of the 21st International Conference on Extending Database Technology}, isbn = {9783893180783}, issn = {2367-2005}, location = {Vienna, Austria}, pages = {145--156}, publisher = {OpenProceedings}, title = {{Synchronous multi-GPU training for deep learning with low-precision communications: An empirical study}}, doi = {10.5441/002/EDBT.2018.14}, year = {2018}, } @inproceedings{7123, abstract = {Population protocols are a popular model of distributed computing, in which n agents with limited local state interact randomly, and cooperate to collectively compute global predicates. Inspired by recent developments in DNA programming, an extensive series of papers, across different communities, has examined the computability and complexity characteristics of this model. Majority, or consensus, is a central task in this model, in which agents need to collectively reach a decision as to which one of two states A or B had a higher initial count. Two metrics are important: the time that a protocol requires to stabilize to an output decision, and the state space size that each agent requires to do so. It is known that majority requires Ω(log log n) states per agent to allow for fast (poly-logarithmic time) stabilization, and that O(log2 n) states are sufficient. Thus, there is an exponential gap between the space upper and lower bounds for this problem. This paper addresses this question. On the negative side, we provide a new lower bound of Ω(log n) states for any protocol which stabilizes in O(n1–c) expected time, for any constant c > 0. This result is conditional on monotonicity and output assumptions, satisfied by all known protocols. Technically, it represents a departure from previous lower bounds, in that it does not rely on the existence of dense configurations. Instead, we introduce a new generalized surgery technique to prove the existence of incorrect executions for any algorithm which would contradict the lower bound. Subsequently, our lower bound also applies to general initial configurations, including ones with a leader. On the positive side, we give a new algorithm for majority which uses O(log n) states, and stabilizes in O(log2 n) expected time. Central to the algorithm is a new leaderless phase clock technique, which allows agents to synchronize in phases of Θ(n log n) consecutive interactions using O(log n) states per agent, exploiting a new connection between population protocols and power-of-two-choices load balancing mechanisms. We also employ our phase clock to build a leader election algorithm with a state space of size O(log n), which stabilizes in O(log2 n) expected time.}, author = {Alistarh, Dan-Adrian and Aspnes, James and Gelashvili, Rati}, booktitle = {Proceedings of the 29th Annual ACM-SIAM Symposium on Discrete Algorithms}, isbn = {9781611975031}, location = {New Orleans, LA, United States}, pages = {2221--2239}, publisher = {ACM}, title = {{Space-optimal majority in population protocols}}, doi = {10.1137/1.9781611975031.144}, year = {2018}, } @inproceedings{397, abstract = {Concurrent sets with range query operations are highly desirable in applications such as in-memory databases. However, few set implementations offer range queries. Known techniques for augmenting data structures with range queries (or operations that can be used to build range queries) have numerous problems that limit their usefulness. For example, they impose high overhead or rely heavily on garbage collection. In this work, we show how to augment data structures with highly efficient range queries, without relying on garbage collection. We identify a property of epoch-based memory reclamation algorithms that makes them ideal for implementing range queries, and produce three algorithms, which use locks, transactional memory and lock-free techniques, respectively. Our algorithms are applicable to more data structures than previous work, and are shown to be highly efficient on a large scale Intel system. }, author = {Arbel Raviv, Maya and Brown, Trevor A}, isbn = {978-1-4503-4982-6}, location = {Vienna, Austria}, number = {1}, pages = {14 -- 27}, publisher = {ACM}, title = {{Harnessing epoch-based reclamation for efficient range queries}}, doi = {10.1145/3178487.3178489}, volume = {53}, year = {2018}, } @article{43, abstract = {The initial amount of pathogens required to start an infection within a susceptible host is called the infective dose and is known to vary to a large extent between different pathogen species. We investigate the hypothesis that the differences in infective doses are explained by the mode of action in the underlying mechanism of pathogenesis: Pathogens with locally acting mechanisms tend to have smaller infective doses than pathogens with distantly acting mechanisms. While empirical evidence tends to support the hypothesis, a formal theoretical explanation has been lacking. We give simple analytical models to gain insight into this phenomenon and also investigate a stochastic, spatially explicit, mechanistic within-host model for toxin-dependent bacterial infections. The model shows that pathogens secreting locally acting toxins have smaller infective doses than pathogens secreting diffusive toxins, as hypothesized. While local pathogenetic mechanisms require smaller infective doses, pathogens with distantly acting toxins tend to spread faster and may cause more damage to the host. The proposed model can serve as a basis for the spatially explicit analysis of various virulence factors also in the context of other problems in infection dynamics.}, author = {Rybicki, Joel and Kisdi, Eva and Anttila, Jani}, issn = {1091-6490}, journal = {Proceedings of the National Academy of Sciences of the United States of America}, number = {42}, pages = {10690 -- 10695}, publisher = {National Academy of Sciences}, title = {{Model of bacterial toxin-dependent pathogenesis explains infective dose}}, doi = {10.1073/pnas.1721061115}, volume = {115}, year = {2018}, } @inproceedings{6589, abstract = {Distributed training of massive machine learning models, in particular deep neural networks, via Stochastic Gradient Descent (SGD) is becoming commonplace. Several families of communication-reduction methods, such as quantization, large-batch methods, and gradient sparsification, have been proposed. To date, gradient sparsification methods--where each node sorts gradients by magnitude, and only communicates a subset of the components, accumulating the rest locally--are known to yield some of the largest practical gains. Such methods can reduce the amount of communication per step by up to \emph{three orders of magnitude}, while preserving model accuracy. Yet, this family of methods currently has no theoretical justification. This is the question we address in this paper. We prove that, under analytic assumptions, sparsifying gradients by magnitude with local error correction provides convergence guarantees, for both convex and non-convex smooth objectives, for data-parallel SGD. The main insight is that sparsification methods implicitly maintain bounds on the maximum impact of stale updates, thanks to selection by magnitude. Our analysis and empirical validation also reveal that these methods do require analytical conditions to converge well, justifying existing heuristics.}, author = {Alistarh, Dan-Adrian and Hoefler, Torsten and Johansson, Mikael and Konstantinov, Nikola H and Khirirat, Sarit and Renggli, Cedric}, booktitle = {32nd Conference on Neural Information Processing Systems}, issn = {1049-5258}, location = {Montreal, Canada}, pages = {5973--5983}, publisher = {Neural Information Processing Systems Foundation}, title = {{The convergence of sparsified gradient methods}}, year = {2018}, } @inproceedings{487, abstract = {In this paper we study network architecture for unlicensed cellular networking for outdoor coverage in TV white spaces. The main technology proposed for TV white spaces is 802.11af, a Wi-Fi variant adapted for TV frequencies. However, 802.11af is originally designed for improved indoor propagation. We show that long links, typical for outdoor use, exacerbate known Wi-Fi issues, such as hidden and exposed terminal, and significantly reduce its efficiency. Instead, we propose CellFi, an alternative architecture based on LTE. LTE is designed for long-range coverage and throughput efficiency, but it is also designed to operate in tightly controlled and centrally managed networks. CellFi overcomes these problems by designing an LTE-compatible spectrum database component, mandatory for TV white space networking, and introducing an interference management component for distributed coordination. CellFi interference management is compatible with existing LTE mechanisms, requires no explicit communication between base stations, and is more efficient than CSMA for long links. We evaluate our design through extensive real world evaluation on of-the-shelf LTE equipment and simulations. We show that, compared to 802.11af, it increases coverage by 40% and reduces median flow completion times by 2.3x.}, author = {Baig, Ghufran and Radunovic, Bozidar and Alistarh, Dan-Adrian and Balkwill, Matthew and Karagiannis, Thomas and Qiu, Lili}, booktitle = {Proceedings of the 2017 13th International Conference on emerging Networking EXperiments and Technologies}, isbn = {978-145035422-6}, location = {Incheon, South Korea}, pages = {2 -- 14}, publisher = {ACM}, title = {{Towards unlicensed cellular networks in TV white spaces}}, doi = {10.1145/3143361.3143367}, year = {2017}, } @inproceedings{791, abstract = {Consider the following random process: we are given n queues, into which elements of increasing labels are inserted uniformly at random. To remove an element, we pick two queues at random, and remove the element of lower label (higher priority) among the two. The cost of a removal is the rank of the label removed, among labels still present in any of the queues, that is, the distance from the optimal choice at each step. Variants of this strategy are prevalent in state-of-the-art concurrent priority queue implementations. Nonetheless, it is not known whether such implementations provide any rank guarantees, even in a sequential model. We answer this question, showing that this strategy provides surprisingly strong guarantees: Although the single-choice process, where we always insert and remove from a single randomly chosen queue, has degrading cost, going to infinity as we increase the number of steps, in the two choice process, the expected rank of a removed element is O(n) while the expected worst-case cost is O(n log n). These bounds are tight, and hold irrespective of the number of steps for which we run the process. The argument is based on a new technical connection between "heavily loaded" balls-into-bins processes and priority scheduling. Our analytic results inspire a new concurrent priority queue implementation, which improves upon the state of the art in terms of practical performance.}, author = {Alistarh, Dan-Adrian and Kopinsky, Justin and Li, Jerry and Nadiradze, Giorgi}, booktitle = {Proceedings of the ACM Symposium on Principles of Distributed Computing}, isbn = {978-145034992-5}, location = {Washington, WA, USA}, pages = {283 -- 292}, publisher = {ACM}, title = {{The power of choice in priority scheduling}}, doi = {10.1145/3087801.3087810}, volume = {Part F129314}, year = {2017}, } @inproceedings{431, abstract = {Parallel implementations of stochastic gradient descent (SGD) have received significant research attention, thanks to its excellent scalability properties. A fundamental barrier when parallelizing SGD is the high bandwidth cost of communicating gradient updates between nodes; consequently, several lossy compresion heuristics have been proposed, by which nodes only communicate quantized gradients. Although effective in practice, these heuristics do not always converge. In this paper, we propose Quantized SGD (QSGD), a family of compression schemes with convergence guarantees and good practical performance. QSGD allows the user to smoothly trade off communication bandwidth and convergence time: nodes can adjust the number of bits sent per iteration, at the cost of possibly higher variance. We show that this trade-off is inherent, in the sense that improving it past some threshold would violate information-theoretic lower bounds. QSGD guarantees convergence for convex and non-convex objectives, under asynchrony, and can be extended to stochastic variance-reduced techniques. When applied to training deep neural networks for image classification and automated speech recognition, QSGD leads to significant reductions in end-to-end training time. For instance, on 16GPUs, we can train the ResNet-152 network to full accuracy on ImageNet 1.8 × faster than the full-precision variant. }, author = {Alistarh, Dan-Adrian and Grubic, Demjan and Li, Jerry and Tomioka, Ryota and Vojnović, Milan}, issn = {1049-5258}, location = {Long Beach, CA, United States}, pages = {1710--1721}, publisher = {Neural Information Processing Systems Foundation}, title = {{QSGD: Communication-efficient SGD via gradient quantization and encoding}}, volume = {2017}, year = {2017}, } @inproceedings{432, abstract = {Recently there has been significant interest in training machine-learning models at low precision: by reducing precision, one can reduce computation and communication by one order of magnitude. We examine training at reduced precision, both from a theoretical and practical perspective, and ask: is it possible to train models at end-to-end low precision with provable guarantees? Can this lead to consistent order-of-magnitude speedups? We mainly focus on linear models, and the answer is yes for linear models. We develop a simple framework called ZipML based on one simple but novel strategy called double sampling. Our ZipML framework is able to execute training at low precision with no bias, guaranteeing convergence, whereas naive quanti- zation would introduce significant bias. We val- idate our framework across a range of applica- tions, and show that it enables an FPGA proto- type that is up to 6.5 × faster than an implemen- tation using full 32-bit precision. We further de- velop a variance-optimal stochastic quantization strategy and show that it can make a significant difference in a variety of settings. When applied to linear models together with double sampling, we save up to another 1.7 × in data movement compared with uniform quantization. When training deep networks with quantized models, we achieve higher accuracy than the state-of-the- art XNOR-Net. }, author = {Zhang, Hantian and Li, Jerry and Kara, Kaan and Alistarh, Dan-Adrian and Liu, Ji and Zhang, Ce}, booktitle = {Proceedings of Machine Learning Research}, isbn = {978-151085514-4}, location = {Sydney, Australia}, pages = {4035 -- 4043}, publisher = {ML Research Press}, title = {{ZipML: Training linear models with end-to-end low precision, and a little bit of deep learning}}, volume = { 70}, year = {2017}, }