DATAMOVE: Mouvements de données pour le calcul haute performance

As digital services are increasingly being deployed and used in a variety of domains, the environmental impact of Information and Communication Technologies (ICTs) is a matter of concern. Artificial intelligence is driving some of this growth but its environmental cost remains scarcely studied. A recent trend in large-scale generative models such as ChatGPT has especially drawn attention since their training requires intensive use of a massive number of specialized computing resources. The inference of those models is made accessible on the web as services, and using them additionally mobilizes end-user terminals, networks, and data centers. Therefore, those services contribute to global warming, worsen metal scarcity, and increase energy consumption. This work proposes an LCA-based methodology for a multi-criteria evaluation of the environmental impact of generative AI services, considering embodied and usage costs of all the resources required for training models, inferring from them, and hosting them online. We illustrate our methodology with Stable Diffusion as a service, an open-source text-to-image generative deep-learning model accessible online. This use case is based on an experimental observation of Stable Diffusion training and inference energy consumption. Through a sensitivity analysis, various scenarios estimating the influence of usage intensity on the impact sources are explored.

Today, most high-performance computing (HPC) platforms have heterogeneous hardware resources (CPUs, GPUs, storage, etc.) A Graphics Processing Unit (GPU) is a parallel computing coprocessor specialized in accelerating vector operations. The prediction of application execution times over these devices is a great challenge and is essential for efficient job scheduling. There are different approaches to do this, such as analytical modeling and machine learning techniques. Analytic predictive models are useful, but require manual inclusion of interactions between architecture and software, and may not capture the complex interactions in GPU architectures. Machine learning techniques can learn to capture these interactions without manual intervention, but may require large training sets. In this paper, we compare three different machine learning approaches: linear regression, support vector machines and random forests with a BSP-based analytical model, to predict the execution time of GPU applications. As input to the machine learning algorithms, we use profiling information from 9 different applications executed over 9 different GPUs. We show that machine learning approaches provide reasonable predictions for different cases. Although the predictions were inferior to the analytical model, they required no detailed knowledge of application code, hardware characteristics or explicit modeling. Consequently, whenever a database with profile information is available or can be generated, machine learning techniques can be useful for deploying automated on-line performance prediction for scheduling applications on heterogeneous architectures containing GPUs.

We present a new approach for scheduling independent tasks on multiple CPUs and multiple GPUs. The tasks are assumed to be parallelizable on CPUs using the moldable model: the final number of cores allotted to a task can be decided and set by the scheduler. More precisely, we design an algorithm aiming at minimizing the makespan-the maximum completion time of all tasks-for this scheduling problem. The proposed algorithm combines a dual approximation scheme with a fast integer linear program (ILP). It determines both the partitioning of the tasks, i.e., whether a task should be mapped to CPUs or a GPU, and the number of CPUs allotted to a moldable task if mapped to the CPUs. A worst-case analysis shows that the algorithm has an approximation ratio of 3/2 + ε. Since the time complexity of the ILP-based algorithm could be non-polynomial, we also present a polynomial-time algorithm with an approximation ratio of 2 + ε. We complement the theoretical analysis of our two novel algorithms with a simulation study. In these simulations, we compare our algorithms to a modified version of the classical HEFT algorithm, which we adapted to handle moldable tasks. The simulation results show that our algorithm with the (3/2 + ε)-approximation ratio produces significantly shorter schedules than the modified HEFT for most of the instances. In addition, our results provide evidence that our ILP-based algorithm can solve larger problem instances in a reasonable amount of time.

Understanding the

The EASY-FCFS heuristic is the basic building block of job scheduling policies in most parallel High Performance Computing platforms. Despite its simplicity, and the guarantee of no job starvation, it could still be improved on a per-system basis. Such tuning is difficult because of non-linearities in the scheduling process. The study conducted in this paper considers an online approach to the automatic tuning of the EASY heuristic for HPC platforms. More precisely, we consider the problem of selecting a reordering policy for the job queue under several feedback modes. We show via a comprehensive experimental validation on actual logs that periodic simulation of historical data can be used to recover existing in-hindsight results that allow to divide the average waiting time by almost 2. This results holds even when the simulator results are noisy. Moreover, we show that good performances can still be obtained without a simulator, under what is called bandit feedback - when we can only observe the performance of the algorithm that was picked on the live system. Indeed, a simple multi-armed bandit algorithm can reduce the average waiting time by 40 percent.

Information and Communications Technologies (ICT) are an increasingly important contributor to the environmental crisis. Computer scientists need tools for measuring the footprint of the code they produce and for optimizing it. Running Average Power Limit (RAPL) is a low-level interface designed by Intel that provides a measure of the energy consumption of a CPU (and more) without the need for additional hardware. Since 2017, it is available on most x86 processors, including AMD processors. More and more people are using RAPL for energy measurement, mostly like a black box without deep knowledge of its behavior. Unfortunately, this causes mistakes when implementing measurement tools. In this article, we propose to come back to the basic mechanisms that allow to use RAPL measurements and present a critical analysis of their operations. In addition to long-established mechanisms, we explore the suitability of the recent eBPF technology (formerly and abbreviation for extended Berkeley Packet Filter) for working with RAPL. We release an implementation in Rust that avoids the pitfalls we detected in existing tools, improving correctness, timing accuracy and performance, with desirable properties for monitoring and profiling parallel applications. We provide an experimental study with multiple benchmarks and processor models to evaluate the efficiency of the various mechanisms and their impact on parallel software. We show that no mechanism provides a significant performance advantage over the others. However, they differ significantly in terms of ease-of-use and resiliency. We believe that this work will help the community to develop correct, resilient and lightweight measurement tools.

Despite the impressive growth and size of super-computers, the computational power they provide still cannot match the demand. Efficient and fair resource allocation is a critical task. Super-computers use Resource and Job Management Systems to schedule applications, which is generally done by relying on generic index policies such as First Come First Served and Shortest Processing time First in combination with Backfilling strategies. Unfortunately, such generic policies often fail to exploit specific characteristics of real workloads. In this work, we focus on improving the performance of online schedulers. We study mixed policies, which are created by combining multiple job characteristics in a weighted linear expression, as opposed to classical pure policies which use only a single characteristic. This larger class of scheduling policies aims at providing more flexibility and adaptability. We use space coverage and black-box optimization techniques to explore this new space of mixed policies and we study how can they adapt to the changes in the workload. We perform an extensive experimental campaign through which we show that (1) even the best pure policy is far from optimal and that (2) using a carefully tuned mixed policy would allow to significantly improve the performance of the system. (3) We also provide empirical evidence that there is no one size fits all policy, by showing that the rapid workload evolution seems to prevent classical online learning algorithms from being effective.

The densest k -subgraph problem is a generalization of the maximum clique problem, in which we are given a graph and a positive integer k , and we search among all the subsets of k vertices of the input graph for the subset which induces the maximum number of edges. densest k -subgraph is a well known optimization problem with various applications as, for example, in the design of public encryption schemes, the evaluation of certain financial derivatives, the identification of communities with similar characteristics, etc. In this paper, we first present algorithms for finding exact solutions for densest k -subgraph which improve upon the standard exponential time complexity of an exhaustive enumeration by creating a link between the computation of an optimum for this problem to the computation of other graph-parameters such as dominating set, vertex cover, longest path, etc. An FPT algorithm is also proposed which considers as a parameter the size of the minimum vertex cover. Finally, we present several approximation algorithms which run in moderately exponential or parameterized time, describing trade-offs between complexity and approximability. In contrast with most of the algorithms in the bibliography, our algorithms need only polynomial space.

With the increase of demand for computing resources and the struggle to provide the necessary energy, power-aware resource management is becoming a major issue for the High-performance computing (HPC) community. Including reliable energy management to a supercomputer’s resource and job management system (RJMS) is not an easy task. The energy consumption of jobs is rarely known in advance and the workload of every machine is unique and different from the others. We argue that the first step toward properly managing energy is to deeply understand the energy consumption of the workload, which involves predicting the workload’s power consumption and exploiting it by using smart power-aware scheduling algorithms. Crucial questions are (i) how sophisticated a prediction method needs to be to provide accurate workload power predictions, and (ii) to what point an accurate workload’s power prediction translates into efficient energy management. In this work, we propose a method to predict and exploit HPC workloads’ power consumption, with the objective of reducing the supercomputer’s power consumption while maintaining the management (scheduling) performance of the RJMS. Our method exploits workload submission logs with power monitoring data, and relies on a mix of light-weight power prediction methods and a classical EASY Backfillling inspired heuristic. We base this study on logs of Marconi 100, a 980 servers supercomputer. We show using simulation that a light-weight history-based prediction method can provide accurate enough power prediction to improve the energy management of a large scale supercomputer compared to energy-unaware scheduling algorithms. These improvements have no significant negative impact on performance.

The in situ paradigm proposes to co-locate simulation and analytics on the same compute node to analyze data while still resident in the compute node memory, hence reducing the need for post-processing methods. A standard approach that proved efficient for sharing resources on each node consists in running the analytics processes on a set of dedicated cores, called helper cores, to isolate them from the simulation processes. Simulation and analytics thus run concurrently with limited interference. In this paper we show that the performance can be improved through a dynamic helper core strategy . We rely on a work stealing scheduler to implement TINS, a task-based in situ framework with an on-demand analytics isolation. The helper cores are dedicated to analytics only when analytics tasks are available. Otherwise the helper cores join the other cores for processing simulation tasks. TINS relies on the Intel ® TBB library. Experiments on up to 14,336 cores run a set of representative analytics parallelized with TBB coupled with the hybrid MPI+TBB ExaStamp molecular dynamics code. TINS shows up to 40% performance improvement over various other approaches including the standard helper core. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

International audience

The growing appetite of new technologies, such as Internet-of-Things, for Cloud resources leads to an unprecedented energy consumption for these infrastructures. In order to make these energy-hungry distributed systems more sustainable, Cloud providers resort more and more to on-site renewable energy production facilities like photovoltaic panels. Yet, this intermittent and variable electricity production is often uncorrelated with the Cloud consumption induced by its workload. Geographical load balancing, virtual machine (VM) migration and consolidation can be used to exploit multiple Cloud data centers' locations and their associated photovoltaic panels for increasing their renewable energy consumption. However, these techniques cost energy and network bandwidth, and this limits their utilization. In this paper, we propose to rely on the flexibility brought by Smart Grids to exchange renewable energy between distributed sites and thus, to further increase the overall Cloud's self-consumption of the locally-produced renewable energy. Our solution is named SCORPIUS: Self-Consumption Optimization of Renewable energy Production In distribUted cloudS. It takes into account telecommunication network constraints and electrical grid requirements to optimize the Cloud's self-consumption by trading-off between VM migration and renewable energy exchange. Our simulation-based results show that SCORPIUS outperforms existing solutions on various workload traces of production Clouds in terms of both renewable self-consumption and overall energy consumption.

We propose heuristics for approximating the maximum cardinality matching on undirected graphs. Our heuristics are based on the theoretical body of a certain type of random graphs, and are made practical for real-life ones. The idea is based on judiciously selecting a subgraph of a given graph and obtaining a maximum cardinality matching on this subgraph. We show that the heuristics have an approximation guarantee of around 0.866 — log(n)/n for a graph with n vertices. Experiments for verifying the theoretical results in practice are provided.

With the increase of demand for computing resources and the struggle to provide the necessary energy, power-aware resource management is becoming a major issue for the High-performance computing (HPC) community. Including reliable energy management to a supercomputer's resource and job management system (RJMS) is not an easy task. The energy consumption of jobs is rarely known in advance and the workload of every machine is unique and different from the others. We argue that the first step towards properly managing power is to deeply understand the power consumption of the workload, which involves predicting the workload power consumption and exploiting it by using smart power-aware scheduling algorithms. Crucial questions are (i) how sophisticated a prediction method needs to be to provide accurate workload power predictions, and (ii) to what point an accurate workload's power prediction translates into efficient power management. In this work, we proposed a method to predict and exploit HPC workloads power consumption with the objective of reducing the supercomputers power consumption, while maintaining the management (scheduling) performance of the RJMS. Our method exploits workload submission logs with power monitoring data, and relies on a mix of lightweight power prediction methods and a classical EASY Backfillling inspired heuristic. Then, we model and solve the power capping scheduling as a greedy knapsack algorithm. This algorithm improves the Quality of Service and avoids starvation while keeping the solution lightweight. We base this study on logs of Marconi 100, a 980-node supercomputer. We show using simulation that a lightweight history-based prediction method can provide accurate enough power prediction to improve the energy management of a large scale supercomputer compared to energy-unaware scheduling algorithms. These improvements have no significant negative impact on performance.

Cloud and HPC (High-Performance Computing) systems have increasingly become more varying in their behavior, in particular in aspects such as performance and power consumption, and the fact that they are becoming less predictable demands more runtime management. In this work, we describe results addressing autonomic administration in HPC systems for scientific workflows management through a control theoretical approach. We propose a model described by parameters related to the key aspects of the infrastructure thus achieving a deterministic dynamical representation that covers the diverse and time-varying behaviors of the real computing system. Later, we propose a model-predictive control loop to achieve two different objectives: maximize cluster utilization by best-effort jobs and control the file server's load in the presence of external disturbances. The accuracy of the prediction relies on a parameter estimation scheme based on the EKF (Extended Kalman Filter) to adjust the predictive-model to the real system, making the approach adaptive to parametric variations in the infrastructure. The closed loop strategy shows performance improvement and consequently a reduction in the total computation time. The problem is addressed in a general way, to allow the implementation on similar HPC platforms, as well as scalability to different infrastructures.

One of the most important challenges for the Internet of Things (IoT) is the implementation of edge computing platforms. In order to improve the response time, the confidentiality, or the energy consumption of IoT applications, part of IoT services must be operated on servers, deployed close to connected devices. This said, much remains to be done for the practical realization of this vision. What are edge servers? Where will they be deployed? These are some questions that are poorly addressed in the literature. In this paper, we propose to build an edge computing framework around new datacenter models in which computing servers are seamlessly integrated in buildings. The datacenter models we are considering have been developed around the concept of data furnace. This concept has been implemented in European cities for district heating. Our paper introduces a new processing model in which edge and distributed cloud computing are operated with the same data furnace servers. We then discuss the challenges in the utilization of this model for edge computing. Our discussion deals with urban integration, service computing frameworks and performance issues. Finally, we end with an impact analysis of our proposition on the future of cloud computing.

Statement of

Distributed Clouds are nowadays an essential component for providing Internet services to always more numerous connected devices. This growth leads the energy consumption of these distributed infrastructures to be a worrying environmental and economic concern. In order to reduce energy costs and carbon footprint, Cloud providers could resort to producing onsite renewable energy, with solar panels for instance. In this paper, we propose NEMESIS: a Network-aware Energy-efficient Management framework for distributEd cloudS Infrastructures with on-Site photovoltaic production. NEMESIS optimizes VM placement and balances VM migration and green energy consumption in Cloud infrastructure embedding geographically distributed data centers with on-site photovoltaic power supply. We use the Simgrid simulation toolbox to evaluate the energy efficiency of NEMESIS against state-of-the-art approaches.

Nowadays, the operations performed by the Internet of Things (IoT) systems are no more trivial since they rely on more sophisticated devices than in the past. The IoT system is physically composed of connected computing, digital, mechanical devices such as sensors or actuators . Most of the time, each of them incorporates a logical arithmetic unit that can pre-compute or compute on the device. To extract value from the data produced at the edge, processing power offered by cloud computing is still utilized. However, streaming data to the cloud exposes some limitations related to the increased communication and data transfer , which introduces delays and consumes network bandwidth . Clustering data is one example of a treatment that can be executed in the cloud. In this paper, we propose a methodology for solving the data stream clustering problem at the edge. Data Stream clustering is defined as the clustering of data that arrive continuously, such as telephone records, multimedia data, sensors data, financial transactions, etc. Since we use low-cost and low-capacity devices, the objective is, given a sequence of points, to construct a good clustering of the stream using a small amount of memory and time. We propose a ‘windowing’ scheme, coupled with a sampling scheme to respect the objective. Under the experimental conditions, experiments show that the clustering solutions can be controlled, with difficulties for time-stamped data but not for random data or data with well-delimited clusters. The main advantage of our schema is that we are clustering data “on the fly” with no knowledge or assumption regarding the available data. We do not assume that all the data are known before a treatment batch by batch. Our schema also has the potential to be adapted to other classes of machine learning algorithms .

International audience