Search

There are massive jobs submitted in the supercomputer, and the job management system is typically deployed to schedule these jobs and allocate compute resources. FCFS (First Come First Serve) is a popular scheduling policy and the job's priority is determined based on the arrival time. However, under the FCFS strategy, if existing idle resources cannot meet the requirement of the head job in the waiting queue, they cannot be allocated to other jobs, which suffers from resource waste. To optimize the resource utilization, the backfilling method is proposed, which allocates the reserved idle compute nodes to a small-size and short-running non-head job, on the premise of not delaying the original head job. Obtaining the job's runtime in advance is necessary for backfilling and the traditional method relies on the user's estimation. Unfortunately, the estimated runtime provided by users is generally overestimated. Many studies extract features from historical job logs and adopt machine learning to predict the runtime. However, traditional features are insufficient to characterize the job. In this paper, we collect job logs from two supercomputers and present a novel runtime prediction framework called PREP. PREP explores the job's running path as a new feature, which implies plentiful information about the job's properties, such as the user, the project, the scale of data sets, and the parameters used. As there is a strong correlation between the job's runtime and its running path, we group jobs with similar paths into a cluster and train a runtime prediction model for each cluster respectively. Extensive evaluations demonstrate that introducing the new feature can achieve higher prediction accuracy (88.5% and 82.3% in two production supercomputers respectively), and our framework has a more desirable prediction performance than other popular strategies like Last-2 and IRPA. In addition, the predicted runtime is inserted into the real job trace of a slurm simulator to verify the advantages of PREP. • PREP explores the job's running path as a new feature to support runtime prediction. • A strong correlation between the job's runtime and its running path is evaluated. • Multiple machine learning algorithms are adopted and compared when predicting. • PREP framework suits different types of supercomputers and gain desirable prediction result. [ABSTRACT FROM AUTHOR]

Communication overhead is the key challenge for distributed training. Gradient compression is a widely used approach to reduce communication traffic. When combined with a parallel communication mechanism method like pipeline, gradient compression technique can greatly alleviate the impact of communication overhead. However, there exist two problems of gradient compression technique to be solved. Firstly, gradient compression brings in extra computation cost, which will delay the next training iteration. Secondly, gradient compression usually leads to a decrease in convergence accuracy. In this paper, we combine parallel mechanism with gradient quantization and periodic full-gradient compensation, and propose a new distributed optimization method named CP-SGD, which can hide the overhead of gradient compression, overlap part of the communication and obtain high convergence accuracy. The local update operation in CP-SGD allows the next iteration to be launched quickly without waiting for the completion of gradient compression and the current communication process. Besides, the accuracy loss caused by gradient compression is solved by k-step correction method introduced in CP-SGD, which provides a gradient correction every k iterations. We prove that CP-SGD has a convergence guarantee and it achieves at least O (1 K + 1 K) convergence rate, where K is the number of iterations. We conduct extensive experiments on MXNet to verify the convergence properties and scaling performance of CP-SGD. Experimental results on a 32-GPU cluster show that convergence accuracy of CP-SGD is close to or even slightly better than that of S-SGD, and its end-to-end time is 30% less than 2-bit gradient compression under a 56Gbps bandwidth environment. In addition, we analyze the performance of CP-SGD when training on 8, 16 and 32 GPUs. It is found that CP-SGD is suitable for most compression-supported update algorithms, and its scalability is approximately linear. • CP-SGD: Distributed Stochastic Gradient Descent with Compression and Periodic Compensation. • CP-SGD help distributed deep learning train with high-accuracy and friendly-communication simultaneously. • Combining the advantages of parallel mechanism and gradient compression. • K-step correction technique makes CP-SGD achieve codesign optimizations and improve compression-based methods greatly. • CP-SGD suits most compression-supported update algorithms and gradient compression strategies. [ABSTRACT FROM AUTHOR]

Decentralized Online Social Network (DOSN) services have been proposed to protect data privacy. In DOSN, the data published by a user and their replicas are only stored in the friend circle of the user. Although full replication can improve Data Availability (DA), pure DOSNs may not deliver sustainable DA. This paper proposes a Cloud-assisted data replication and storage scheme, called Cadros, to improve the DA in DOSN. This paper conducts quantitative analysis about the storage capacity of Cadros, and further models and predicts the level of DA that Cadros can achieve. The data in Cadros are partitioned in such a way that the overhead caused by storing the data in the Cloud is minimized while satisfying the desired DA. This paper also proposes the data placement strategies to realize the desired DA and improve other performance. Experiments have been conducted to verify the effectiveness of Cadros. [ABSTRACT FROM AUTHOR]