New paper “Why Globally Re-shuffle? Revisiting Data Shuffling in Large Scale Deep Learning” to be presented at IPDPS’22.
New paper “Why Globally Re-shuffle? Revisiting Data Shuffling in Large Scale Deep Learning” to be presented at IPDPS’22.
Available online: https://hal.archives-ouvertes.fr/hal-03599740/document
Abstract
Stochastic gradient descent (SGD) is the most prevalent algorithm for training Deep Neural Networks (DNN). SGD iterates the input data set in each training epoch processing data samples in a random access fashion. Because this puts enormous pressure on the I/O subsystem, the most common approach to distributed SGD in HPC environments is to replicate the entire dataset to node local SSDs. However, due to rapidly growing data set sizes this approach has become increasingly infeasible. Surprisingly, the questions of why and to what extent random access is required have not received a lot of attention in the literature from an empirical standpoint.
In this paper, we revisit data shuffling in DL workloads to investigate the viability of partitioning the dataset among workers and performing only a partial distributed exchange of samples in each training epoch. Through extensive experiments on up to 2,048 GPUs of ABCI and 4,096 compute nodes of Fugaku, we demonstrate that in practice validation accuracy of global shuffling can be maintained when carefully tuning the partial distributed exchange. We provide a solution implemented in PyTorch that enables users to control the proposed data exchange scheme.
New paper “Blockchain logging for process mining: a systematic review” at HICSS’2022
Authors: Leyla Moctar M’Baba, Mohamed Sellami, Walid Gaaloul and Mohamedade Farouk NANNE
Abstract
Considerable progress was forcasted for collaborative business processes with the rise of blockchain programmable platforms. One of the saliant promises was auditable traces of business process execution, but practically it has posed challenges specially with regard to blockchain logs’ structure who turned out to be inadequate for process mining techniques. Approaches to answer this issue have started to emerge in the literature; some focusing on the creation process of event logs, and others dealing with their retrieval from the blockchain. This work outlines the generic steps required to solve these challenges and analyzes findings in these approaches with a consideration for efficiency and future research directions.
New paper “Towards higher-level description of SLA-aware reconfiguration strategies based on state-machine” at ICEBE’2021
Authors: Jeremy Mechouche, Roua Touihri, Mohamed Sellami and Walid Gaaloul
Abstract
High number of European projects and international initiatives show an increased interest in the multi-cloud paradigm. One key need identified in these studies is an SLA-driven service model for multi-cloud environment. While offering a multi-cloud application, cloud consumer define reconfiguration strategies to avoid violating SLAs established with their customers. In this context, this paper presents an approach for enriching multi-cloud SLA representations with reconfiguration strategies. Advantages of this approach are twofold: (i) simplify SLA administration and (ii) limit SLA violations caused by reconfiguration strategies. We represent reconfiguration strategies based on state-machine formalism. Furthermore, we define thresholds to guarantee their compliance with multi-cloud SLAs and anticipate SLA violations. An implementation of the approach is presented in the paper and illustrates how these thresholds are computed.
New paper “Runtime models and evolution graphs for the version management of microservice architectures” at APSEC 2021
Authors: Yuwei Wang, Denis Conan, Sophie Chabridon, Kavoos Bojnourdi, Jingxuan Ma.
APSEC 2021, https://hal.archives-ouvertes.fr/hal-03419462
Abstract
Microservice architectures focus on developing modular and independent functional units, which can be automatically deployed, enabling agile DevOps. One major challenge is to manage the rapid evolutionary changes in microservices and perform continuous redeployment without interrupting the application execution. The existing solutions provide limited capacities to help software architects model, plan, and perform version management activities. The architects lack a representation of a microservice architecture with versions tracking. In this paper, we propose runtime models that distinguishes the type model from the instance model, and we build up an evolution graph of configuration snapshots of types and instances to allow the traceability of microservice versions and their deployment. We demonstrate our solution with an illustrative application that involves synchronous (RPC calls) and asynchronous (publish-subscribe) interaction within information systems.
New paper “J-NVM: Off-heap Persistent Objects in Java” to be presented at SOSP’21
New paper “J-NVM: Off-heap Persistent Objects in Java” to be presented at SOSP’21. Congrats to Anatole, Yohan, Kwabena, Pierre and Gaël!
New paper “Montsalvat: Intel SGX Shielding for GraalVM Native Images” to be presented at Middleware’21
New paper “Montsalvat: Intel SGX Shielding for GraalVM Native Images” to be presented at Middleware’21. Congrats to Gaël!
New paper “The Serverless Shell” to be presented at Middleware’21
New paper “The Serverless Shell” to be presented at Middleware’21. Congrats to Aurele and Pierre!
New paper “Highly-available and consistent group collaboration at the edge with Colony” to be presented at Middleware’21
New paper “Highly-available and consistent group collaboration at the edge with Colony” to be presented at Middleware’21. Congrats to Pierre!
New paper “Automating user-feedback driven requirements prioritization” in Elsevier Information and Software Technology
Authors: Fitsum Meshesha Kifetew, Anna Perini, Angelo Susi, Aberto Siena, Denisse Muñante and Itzel Morales-Ramirez
Information and Software Technology, Elsevier, 2021, 138, https://hal.archives-ouvertes.fr/hal-03277970
Abstract
Context: Feedback from end users of software applications is a valuable resource in understanding what users request, what they value, and what they dislike. Information derived from user-feedback can support software evolution activities, such as requirements prioritization. User-feedback analysis is still mostly performed manually by practitioners, despite growing research in automated analysis. Objective: We address two issues in automated user-feedback analysis: (i) most of the existing automated analysis approaches that exploit linguistic analysis assume that the vocabulary adopted by users (when expressing feedback) and developers (when formulating requirements) are the same; and (ii) user-feedback analysis techniques are usually experimentally evaluated only on some user-feedback dataset, not involving assessment by potential software developers. Method: We propose an approach, ReFeed, that computes, for each requirement, the set of related user-feedback, and from such user-feedback extracts quantifiable properties which are relevant for prioritizing the requirement. The extracted properties are propagated to the related requirements, based on which ranks are computed for each requirement. ReFeed relies on domain knowledge, in the form of an ontology, helping mitigate the gap in the vocabulary of end users and developers. The effectiveness of ReFeed is evaluated on a realistic requirements prioritization scenario in two experiments involving graduate students from two different universities. Results: ReFeed is able to synthesize reasonable priorities for a given set of requirements based on properties derived from user-feedback. The implementation of ReFeed and related resources are publicly available. Conclusion: The results from our studies are encouraging in that using only three properties of user-feedback, ReFeed is able to prioritize requirements with reasonable accuracy. Such automatically determined prioritization could serve as a good starting point for requirements experts involved in the task of prioritizing requirements Future studies could explore additional user-feedback properties to improve the effectiveness of computed priorities.
New paper “PrioDeX: a Data Exchange middleware for efficient event prioritization in SDN-based IoT systems” in ACM TOIT
Authors: Georgios Bouloukakis, Kyle Benson, Luca Scalzotto, Paolo Bellavista, Casey Grant, Valérie Issarny, Sharad Mehrotra,Ioannis Moscholios, Nalini Venkatasubramanian
ACM Transactions on Internet of Things, In press, https://hal.archives-ouvertes.fr/hal-03171358
Abstract
Real-time event detection and targeted decision making for emerging mission-critical applications require systems that extract and process relevant data from IoT sources in smart spaces. Oftentimes, this data is heterogeneous in size, relevance, and urgency, which creates a challenge when considering that different groups of stakeholders (e.g., first responders, medical staff, government officials, etc) require such data to be delivered in a reliable and timely manner. Furthermore, in mission-critical settings, networks can become constrained due to lossy channels and failed components, which ultimately add to the complexity of the problem. In this paper, we propose PrioDeX, a cross-layer middleware system that enables timely and reliable delivery of mission-critical data from IoT sources to relevant consumers through the prioritization of messages. It integrates parameters at the application, network, and middleware layers into a data exchange service that accurately estimates end-to-end performance metrics through a queueing analytical model. PrioDeX proposes novel algorithms that utilize the results of this analysis to tune data exchange configurations (event priorities and dropping policies), which is necessary for satisfying situational awareness requirements and resource constraints. PrioDeX leverages Software-Defined Networking (SDN) methodologies to enforce these configurations in the IoT network infrastructure. We evaluate our approach using both simulated and prototype-based experiments in a smart building fire response scenario. Our application-aware prioritization algorithm improves the value of exchanged information by 36% when compared with no prioritization; the addition of our network-aware drop rate policies improves this performance by 42% over priorities only and by 94% over no prioritization.
