In deep learning(DL) training, unique IO characteristics are observed. Although many accelerators (GPUs, FPGAs) exist for DL training, these devices are primarily used for neural network computation graphs. For some IO-intensive DL training tasks, IO becomes a bottleneck. Data augmentation is a data processing method, which helps to improve the diversity of data and the generalization of the model, it creates tremendous repeated computations. At the same time, DL frameworks treat each sample equally, even though the sample contribution to the model varis during training. Existing training frameworks do not leverage these features to build cache layers that reduce reading samples from low-speed storage devices. To address these challenges, we propose ACache, a caching system built using importance sampling. The workload caused by data augmentation is optimized to minimize the interference of repeated computations on training. ACache will update the sampling factors periodically, which are updated asynchronously throughout the training process. Additionally, it implements a fine-grained cache replacement strategy to determine whether to replace the cache for each sample. We evaluated ACache on a representative computer vision task and found that, compared to the current state-of-the-art DL Cache system, ACache is capable of training models at a faster rate.

Existing Text-Image Person Retrieval (TIPR) methods have made substantial progress in modeling cross-modal associations via contrastive learning frameworks, but usually ignore the fine-grained differences in semantic relevance among different samples, which limits retrieval accuracy. To address this problem, we propose a novel Hierarchical Cross-modal Association framework HCA, which leverages the intra-modal fine-grained semantic relations distilled by single-modal pretrained models to constrain hierarchical cross-modal association between image and text modalities, for accurate TIPR. Specifically, to model hierarchical cross-modal semantic relationships, we propose a Hierarchical Relevance Matching (HRM) module. It partitions the matching strength of image-text pairs by jointly considering identity labels and cross-modal similarity, collaborating with unimodal similarity to construct a hierarchical relevance distribution that serves as a soft supervision signal. HRM not only helps the model better capture varying levels of semantic consistency between image-text pairs but also enhances the overall accuracy of cross-modal association learning. To enhance the ability to capture fine-grained cross-modal semantic relationships, we introduce an Image-guided Ambiguous text Token Modeling (IATM) module. It replaces original tokens with semantically ambiguous ones and leverages image guidance to detect and correct these tokens. This process further improves the fine-grained semantic alignment between images and texts. Experimental results demonstrate that HCA achieves new state-of-the-art performance across multiple datasets, thoroughly validating its effectiveness and advancement in cross-modal retrieval tasks.

Transformer-based RGBT tracking has attracted much attention due to the strong modeling capacity of self attention and cross attention mechanisms. These attention mechanisms utilize the correlations among tokens to construct powerful feature representations, but are easily affected by low-quality tokens. To address this issue, we propose a novel Quality-aware Spatio-temporal Transformer Network (QSTNet), which calculates the quality weights of tokens in search regions based on the correlation with multimodal template tokens to suppress the negative effects of low-quality tokens in spatio-temporal feature representations, for robust RGBT tracking. In particular, we argue that the correlation between search tokens of one modality and multimodal template tokens could reflect the quality of these search tokens, and thus design the Quality-aware Token Weighting Module (QTWM) based on the correlation matrix of search and template tokens to suppress the negative effects of low-quality tokens. Specifically, we calculate the difference matrix derived from the attention matrices of the search tokens from both modalities and the multimodal template tokens, and then assign the quality weight for each search token based on the difference matrix, which reflects the relative correlation of search tokens from different modalities to multimodal template tokens. In addition, we propose the Prompt-based Spatio-temporal Encoder Module (PSEM) to utilize spatio-temporal multimodal information while alleviating the impact of low-quality spatio-temporal features. Extensive experiments on four RGBT benchmark datasets demonstrate that the proposed QSTNet exhibits superior performance compared to other state-of-the-art tracking methods. Our code and supplementary video are now available: https://zhaodongah.github.io/QSTNet.

In recent years, the number of clusters and cloud platforms dedicated to deep learning acceleration has increased, and research on multi-tenant deep learning (DL) cluster scheduling systems has also advanced quickly. However, we have observed several shortcomings in these systems. Firstly, resources exhibit heterogeneity, but even the most advanced heterogeneity-aware schedulers can only reach the GPU-type level. In addition, most scheduling systems cannot perform well in balancing efficiency and fairness, which leads to unfair resource allocation and reduced user satisfaction. Moreover, we have noticed the phenomenon of cluster fragmentation and job starvation.

Deep learning has become a cornerstone in artificial intelligence, prompting significant investment in large-scale model research by businesses, universities, and research institutions. As a result, the reliance on GPU clusters for deep learning research and development has intensified, underscoring the importance of efficiently allocating cluster resources. While recent advancements in deep neural network (DNN) training schedulers have improved various performance metrics, they fall short in strategies tailored for large-model training, predominantly focusing on data parallelism and overlooking the crucial role of model parallelism for large-scale models. Addressing this gap, we introduce "Morph", a novel approach that predicts throughput for model parallel tasks by evaluating all possible resource allocations. Morph leverages queueing theory to assess cluster load, optimizing resource distribution based on the cluster’s processing capabilities. Additionally, it employs a simulated annealing algorithm to refine the sequence of task queues, significantly enhancing cluster throughput. Our experiments demonstrate that Morph, integrated with existing schedulers, achieves more effective resource allocation, optimizes task sequences, and improves cluster throughput, markedly reducing average job completion time by up to 50.69% and queue waiting times by up to 71.01%.

Traditional persona-based dialog generation models typically rely on textual descriptions of persona information, but this approach is limited by this and its inability to capture all aspects of a persona’s speaking style. To solve this problem, we propose a novel dialog generation model called Speaker-History-Aware Dialogue GEneration (SHADE) through Contrastive and Prompt Learning, which utilizes contrastive learning to model speaking style from historical conversation, resulting in more personalized and distinctive response. Based on the idea of, Prompt learning, on the one hand, we embed small-scale trainable parameters according to Continuous Prompts to stimulate the knowledge of the pre-trained language model, so as to improve the feature extraction ability of the model. On the other hand, we add speaker tag to the decoder input according to Discrete Prompts to improve the consistency of speaking style. We evaluate our model on two popular text-only datasets, DailyDialog and PersonaChat. The results show that SHADE outperforms other multiturn hierarchical dialog generation models in terms of response consistency, logic, and diversity. SHADE is optimal for metrics PPL, NIST, ROUGEL, METEOR, BLUE2, etc. Our proposed SHADE offers a promising direction for persona-based dialog generation, addressing the limitation of existing approaches and paving the way for more personalized and engaging conversational agents in various applications.

Existing RGBT tracking methods usually localize a target object with a bounding box, in which the trackers are often affected by the inclusion of background clutter. To address this issue, this article presents a novel algorithm, called noise-robust cross-modal ranking, to suppress background effects in target bounding boxes for RGBT tracking. In particular, we handle the noise interference in cross-modal fusion and seed labels from the following two aspects. First, the soft cross-modality consistency is proposed to allow the sparse inconsistency in fusing different modalities, aiming to take both collaboration and heterogeneity of different modalities into account for more effective fusion. Second, the optimal seed learning is designed to handle label noises of ranking seeds caused by some problems, such as irregular object shape and occlusion. In addition, to deploy the complementarity and maintain the structural information of different features within each modality, we perform an individual ranking for each feature and employ a cross-feature consistency to pursue their collaboration. A unified optimization framework with an efficient convergence speed is developed to solve the proposed model. Extensive experiments demonstrate the effectiveness and efficiency of the proposed approach comparing with state-of-the-art tracking methods on GTOT and RGBT234 benchmark data sets.

Context: Design pattern detection is a very important research on software reuse, which can greatly help software maintenance and reconstruction. At present, many researchers have invested in this work and proposed a variety of methods for detection. Method: This paper extends the graph matching technology based on the past and proposes a new method that combines network analysis and structural matching for detection. First, we use network level analysis to obtain important nodes, then use the neighborhood path matching algorithm to match the pattern instance. Result: We describe the detection of five patterns on four open source systems, then analyze and compare with the other three methods, for achieving high precision and recall, which demonstrates that our method is effective. Conclusion: Using combining network analysis with structural matching can well detect these pattern instances, and these instances are also especially important for future software refactoring.

Cloud computing can provide a more cost-effective way to deploy scientific workflows than traditional distributed computing environments such as cluster and grid. Due to the large size of scientific datasets, data placement plays an important role in scientific cloud workflow systems for improving system performance and reducing data transfer cost. Traditional task-level data placement strategy only considers shared datasets within individual workflows to reduce data transfer cost. However, it is obvious that task-level strategy is not necessarily good enough for the situation of multiple workflows at the workflow level. In this paper, a novel workflow-level data placement model is constructed, which regards multiple workflows as a whole. Then, a two-stage data placement strategy is proposed which first pre-allocates initial datasets to proper datacenters during workflow build-time stage, and then dynamically distributes newly generated datasets to appropriate datacenters during runtime stage. Both stages use an efficient discrete particle swarm optimization algorithm to place flexible-location datasets. Comprehensive experiments demonstrate that our workflow-level data placement strategy can be more cost-effective than its task-level counterpart for data-sharing scientific cloud workflows.

Dn-time completion is one of critical QoS (Quality of Service) measurements for massive time-constrained business processes executing in the cloud environment. However, temporal violations inevitably occur during the execution of business workflows due to the uncertainty and dynamic nature of cloud environment. To realize the goal of the target on-time completion rate, workflow temporal verification is employed for monitoring the execution time of workflow activities and handling time delays before the deadline. While current studies on business cloud workflows temporal verification mainly monitor the execution of workflows with fixed observation time intervals along the system timeline, which can result in huge monitoring cost. To reduce the monitoring cost, this paper presents a sliding-window based dynamic temporal checkpoint selection strategy using propagation-aware throughput temporal consistency model for parallel cloud business workflows. The strategy adjusts the next observation time interval according to the temporal verification result at the previous checkpoint. Experimental results show that our strategy can select fewer temporal checkpoints and violation handling points compared with conventional strategies under the same situations.

Cloud computing is an emerging business computing model whose basic idea is to transmit all kinds of resources through the Internet such as storage resources, computing resources, bandwidth and so forth. So users do not need to purchase a large of computing systems to manage their business, on the contrary they only need to purchase the resources according to their needs in order to reduce the cost greatly. Due to the flexibility, convenience and low maintenance cost of cloud computing, more and more service providers choose to deploy their services to the cloud. However, under the influence of uncertain factors such as manufacturing technology, the price and delivery time of virtual machine also been changed. Because of these variations, it is difficult for users to select the appropriate cloud servers at a minimum cost, which will lead to a decline in user experience. To solve the problems above, this paper analyzes the performance of different Ali Cloud servers and proposes a server selection strategy according to the size of the whole instance and the required time of completing the workflow, so as to make the whole workflow instances execution costs as little as possible in the regulations. We also provide experimental results to demonstrate the effectiveness of our selection strategies.

Abstract Workflow temporal verification is conducted to guarantee on‐time completion, which is one of the most important QoS (Quality of Service) dimensions for business processes running in the cloud. However, as today's business systems often need to handle a large number of concurrent customer requests, conventional response‐time based process monitoring strategies conducted in a one‐by‐one fashion cannot be applied efficiently to a large batch of parallel processes because of significant time overhead. Similar situations may also exist in software companies where multiple software projects are carried out at the same time by software developers. To address such a problem, based on a novel runtime throughput consistency model, this paper proposes a QoS‐aware throughput based checkpoint selection strategy, which can dynamically select a small number of checkpoints along the system timeline to facilitate the temporal verification of throughput constraints and achieve the target on‐time completion rate. Experimental results demonstrate that our strategy can achieve the best efficiency and effectiveness compared with the state‐of‐the‐art as and other representative response‐time based checkpoint selection strategies. Copyright © 2016 John Wiley & Sons, Ltd.

Due to fast development of e-commerce, logistics, network and cloud computing, many businesses have changed their traditional production and sale patterns. This brings big opportunities and challenges to logistics. Among the process of logistics, delivery is a key issue. This paper mainly focuses on logistics scheduling based on cloud business workflows. A constrained Dijkstra algorithm is proposed to select an optimal route for logistics transportation with various parameters. We then put forward a distribution logistics scheduling algorithm to solve tracking large numbers of expresses based on business workflows. The simulation indicates the effectiveness of our two novel algorithms.

On-time completion is one of the most important QoS (Quality of Service) dimensions for business processes running in the cloud. While today’s business systems often need to handle thousands of concurrent user requests, process monitoring is basically conducted in a one by one fashion. It is possible to repeat the strategies for monitoring a single process a thousand times to monitor a thousand parallel processes. However, the time overhead will be a thousand-fold increase as well, which brings a big challenge for process monitoring. In this paper, based on a novel runtime throughput consistency model, we propose a QoS-aware throughput based checkpoint selection strategy which can dynamically select a small number of checkpoints along the system timeline to facilitate the temporal verification of throughput constraints and achieve the target on-time completion rate. The experimental results demonstrate that our strategy can achieve the best efficiency and effectiveness compared with the state-of-the-art as well as other representative response-time based checkpoint selection strategies.