As numerous machine learning and other algorithms increase in complexity and data requirements, distributed computing becomes necessary to satisfy the growing computational and storage demands, because it enables parallel execution of smaller tasks that make up a large computing job. However, random fluctuations in task service times lead to straggling tasks with long execution times. Redundancy, in the form of task replication and erasure coding, provides diversity that allows a job to be completed when only a subset of redundant tasks is executed, thus removing the dependency on the straggling tasks. In situations of constrained resources (here a fixed number of parallel servers), increasing redundancy reduces the available resources for parallelism. In this paper, we characterize the diversity vs. parallelism trade-off and identify the optimal strategy, among replication, coding and splitting, which minimizes the expected job completion time. We consider three common service time distributions and establish three models that describe scaling of these distributions with the task size. We find that different distributions with different scaling models operate optimally at different levels of redundancy, and thus may require very different code rates.
相關內容
Models of parallel processing systems typically assume that one has $l$ workers and jobs are split into an equal number of $k=l$ tasks. Splitting jobs into $k > l$ smaller tasks, i.e. using ``tiny tasks'', can yield performance and stability improvements because it reduces the variance in the amount of work assigned to each worker, but as $k$ increases, the overhead involved in scheduling and managing the tasks begins to overtake the performance benefit. We perform extensive experiments on the effects of task granularity on an Apache Spark cluster, and based on these, developed a four-parameter model for task and job overhead that, in simulation, produces sojourn time distributions that match those of the real system. We also present analytical results which illustrate how using tiny tasks improves the stability region of split-merge systems, and analytical bounds on the sojourn and waiting time distributions of both split-merge and single-queue fork-join systems with tiny tasks. Finally we combine the overhead model with the analytical models to produce an analytical approximation to the sojourn and waiting time distributions of systems with tiny tasks which include overhead. Though no longer strict analytical bounds, these approximations matched the Spark experimental results very well in both the split-merge and fork-join cases.
Analytical performance models are very effective in ensuring the quality of service and cost of service deployment remain desirable under different conditions and workloads. While various analytical performance models have been proposed for previous paradigms in cloud computing, serverless computing lacks such models that can provide developers with performance guarantees. Besides, most serverless computing platforms still require developers' input to specify the configuration for their deployment that could affect both the performance and cost of their deployment, without providing them with any direct and immediate feedback. In previous studies, we built such performance models for steady-state and transient analysis of scale-per-request serverless computing platforms (e.g., AWS Lambda, Azure Functions, Google Cloud Functions) that could give developers immediate feedback about the quality of service and cost of their deployments. In this work, we aim to develop analytical performance models for the latest trend in serverless computing platforms that use concurrency value and the rate of requests per second for autoscaling decisions. Examples of such serverless computing platforms are Knative and Google Cloud Run (a managed Knative service by Google). The proposed performance model can help developers and providers predict the performance and cost of deployments with different configurations which could help them tune the configuration toward the best outcome. We validate the applicability and accuracy of the proposed performance model by extensive real-world experimentation on Knative and show that our performance model is able to accurately predict the steady-state characteristics of a given workload with minimal amount of data collection.
We present a general framework for designing approximately revenue-optimal mechanisms for multi-item additive auctions, which applies to both truthful and non-truthful auctions. Given a (not necessarily truthful) single-item auction format $A$ satisfying certain technical conditions, we run simultaneous item auctions augmented with a personalized entry fee for each bidder that must be paid before the auction can be accessed. These entry fees depend only on the prior distribution of bidder types, and in particular are independent of realized bids. We bound the revenue of the resulting two-part tariff mechanism using a novel geometric technique that enables revenue guarantees for many common non-truthful auctions that previously had none. Our approach adapts and extends the duality framework of Cai et al [CDW16] beyond truthful auctions. Our framework can be used with many common auction formats, such as simultaneous first-price, simultaneous second-price, and simultaneous all-pay auctions. Our results for first price and all-pay are the first revenue guarantees of non-truthful mechanisms in multi-dimensional environments, addressing an open question in the literature [RST17]. If all-pay auctions are used, we prove that the resulting mechanism is also credible in the sense that the auctioneer cannot benefit by deviating from the stated mechanism after observing agent bids. This is the first static credible mechanism for multi-item additive auctions that achieves a constant factor of the optimal revenue. If second-price auctions are used, we obtain a truthful $O(1)$-approximate mechanism with fixed entry fees that are amenable to tuning via online learning techniques.
For mitigating Byzantine behaviors in federated learning (FL), most state-of-the-art approaches, such as Bulyan, tend to leverage the similarity of updates from the benign clients. However, in federated learning (FL), data distribution across clients is typically heterogeneous. This makes Byzantine fault mitigation a very challenging task, as even the updates from the benign clients are quite dissimilar from each other. Hence, most prior methods, which treat any update that differs from the majority of other updates as a Byzantine update, exhibit poor performance. We propose DiverseFL, in which rather than comparing each client's update with other updates to detect Byzantine clients, the FL server compares each client's update with a guiding update of that client. Any client whose update diverges from its associated guiding update is tagged as a Byzantine node. The FL server computes the guiding update for each participating client over a small sample of the client's local data that is received only once before training. For preserving the privacy of the shared samples, DiverseFL creates a Trusted Execution Environment (TEE)-based secure enclave within the FL server to receive each client's samples, compute guiding updates, and perform secure aggregation for global model update. In experiments, DiverseFL achieves improvements of up to ~16% in absolute test accuracy over prior benchmarks, and consistently performs closely to OracleSGD, where the server only aggregates the updates from the benign clients. We also analyze convergence rate of DiverseFL with non-IID data, under simplifying assumptions such as strong convexity of local loss.
In this work, we consider a remote monitoring scenario in which multiple sensors share a wireless channel to deliver their status updates to a process monitor via an access point (AP). Moreover, we consider that the sensors randomly arrive and depart from the network as they become active and inactive. The goal of the sensors is to devise a medium access strategy to collectively minimize the long-term mean network \ac{AoI} of their respective processes at the remote monitor. For this purpose, we propose specific modifications to ALOHA-QT algorithm, a distributed medium access algorithm that employs a policy tree (PT) and reinforcement learning (RL) to achieve high throughput. We provide the upper bound on the mean network Age of Information (AoI) for the proposed algorithm along with pointers for selecting its key parameter. The results reveal that the proposed algorithm reduces mean network \ac{AoI} by more than 50 percent for state of the art stationary randomized policies while successfully adjusting to a changing number of active users in the network. The algorithm needs less memory and computation than ALOHA-QT while performing better in terms of AoI.
Neural networks of ads systems usually take input from multiple resources, e.g., query-ad relevance, ad features and user portraits. These inputs are encoded into one-hot or multi-hot binary features, with typically only a tiny fraction of nonzero feature values per example. Deep learning models in online advertising industries can have terabyte-scale parameters that do not fit in the GPU memory nor the CPU main memory on a computing node. For example, a sponsored online advertising system can contain more than $10^{11}$ sparse features, making the neural network a massive model with around 10 TB parameters. In this paper, we introduce a distributed GPU hierarchical parameter server for massive scale deep learning ads systems. We propose a hierarchical workflow that utilizes GPU High-Bandwidth Memory, CPU main memory and SSD as 3-layer hierarchical storage. All the neural network training computations are contained in GPUs. Extensive experiments on real-world data confirm the effectiveness and the scalability of the proposed system. A 4-node hierarchical GPU parameter server can train a model more than 2X faster than a 150-node in-memory distributed parameter server in an MPI cluster. In addition, the price-performance ratio of our proposed system is 4-9 times better than an MPI-cluster solution.
For neural networks (NNs) with rectified linear unit (ReLU) or binary activation functions, we show that their training can be accomplished in a reduced parameter space. Specifically, the weights in each neuron can be trained on the unit sphere, as opposed to the entire space, and the threshold can be trained in a bounded interval, as opposed to the real line. We show that the NNs in the reduced parameter space are mathematically equivalent to the standard NNs with parameters in the whole space. The reduced parameter space shall facilitate the optimization procedure for the network training, as the search space becomes (much) smaller. We demonstrate the improved training performance using numerical examples.
Existing Deep Learning frameworks exclusively use either Parameter Server(PS) approach or MPI parallelism. In this paper, we discuss the drawbacks of such approaches and propose a generic framework supporting both PS and MPI programming paradigms, co-existing at the same time. The key advantage of the new model is to embed the scaling benefits of MPI parallelism into the loosely coupled PS task model. Apart from providing a practical usage model of MPI in cloud, such framework allows for novel communication avoiding algorithms that do parameter averaging in Stochastic Gradient Descent(SGD) approaches. We show how MPI and PS models can synergestically apply algorithms such as Elastic SGD to improve the rate of convergence against existing approaches. These new algorithms directly help scaling SGD clusterwide. Further, we also optimize the critical component of the framework, namely global aggregation or allreduce using a novel concept of tensor collectives. These treat a group of vectors on a node as a single object allowing for the existing single vector algorithms to be directly applicable. We back our claims with sufficient emperical evidence using large scale ImageNet 1K data. Our framework is built upon MXNET but the design is generic and can be adapted to other popular DL infrastructures.
Knowledge graph embedding aims to embed entities and relations of knowledge graphs into low-dimensional vector spaces. Translating embedding methods regard relations as the translation from head entities to tail entities, which achieve the state-of-the-art results among knowledge graph embedding methods. However, a major limitation of these methods is the time consuming training process, which may take several days or even weeks for large knowledge graphs, and result in great difficulty in practical applications. In this paper, we propose an efficient parallel framework for translating embedding methods, called ParTrans-X, which enables the methods to be paralleled without locks by utilizing the distinguished structures of knowledge graphs. Experiments on two datasets with three typical translating embedding methods, i.e., TransE [3], TransH [17], and a more efficient variant TransE- AdaGrad [10] validate that ParTrans-X can speed up the training process by more than an order of magnitude.
In this paper, we study the optimal convergence rate for distributed convex optimization problems in networks. We model the communication restrictions imposed by the network as a set of affine constraints and provide optimal complexity bounds for four different setups, namely: the function $F(\xb) \triangleq \sum_{i=1}^{m}f_i(\xb)$ is strongly convex and smooth, either strongly convex or smooth or just convex. Our results show that Nesterov's accelerated gradient descent on the dual problem can be executed in a distributed manner and obtains the same optimal rates as in the centralized version of the problem (up to constant or logarithmic factors) with an additional cost related to the spectral gap of the interaction matrix. Finally, we discuss some extensions to the proposed setup such as proximal friendly functions, time-varying graphs, improvement of the condition numbers.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191