Section: New Results

Resource Management

Participants : Mohamed Abderrahim, Ronan-Alexandre Cherrueau, Bastien Confais, Jad Darrous, Shadi Ibrahim, Adrien Lebre, Matthieu Simonin, Emile Cadorel, Hélène Coullon, Jean-Marc Menaud.

Our contributions regarding resource management can be divided into two main topics described below: contributions related to (i) geo-distributed cloud infrastructures (e.g., Fog and Edge computing) and (ii) the convergence of Cloud and HPC infrastructures.

Geo-distributed Infrastructures

In [15], we provide reflections regarding how fog/edg infrastructures can be operated. While it is clear that edge infrastructures are required for emerging use-cases related to IoT, VR or NFV, there is currently no resource management system able to deliver all features for the edge that made cloud computing successful (e.g., an OpenStack for the edge). Since building a system from scratch is seen by many as impractical, our community should investigate different appraoches. This study, which has been achieved with Ericsson colleagues, provides a list of the features required to operate and use edge computing resources, and investigate how an existing IaaS manager (i.e., OpenStack) satisfies these requirements. Finally, we identify from this study two approaches to design an edge infrastructure manager that fulfils our requirements, and discuss their pros and cons.

In [18], we propose a new novel VMI management system for distributed cloud infrastructures. Most large cloud providers, like Amazon and Microsoft, replicate their Virtual Machine Images (VMIs) on multiple geographically distributed data centers to offer fast service provisioning. Provisioning a service may require to transfer a VMI over the wide-area network (WAN) and therefore is dictated by the distribution of VMIs and the network bandwidth in-between sites. Nevertheless, existing methods to facilitate VMI management (ie, retrieving VMIs) overlook network heterogeneity in geo-distributed clouds. To deal with such a limitation, we design, implement and evaluate Nitro, a novel VMI management system that helps to minimize the transfer time of VMIs over a heterogeneous WAN. To achieve this goal, Nitro incorporates two complementary features. First, it makes use of deduplication to reduce the amount of data which will be transferred due to the high similarities within an image and in-between images. Second, Nitro is equipped with a network-aware data transfer strategy to effectively exploit links with high bandwidth when acquiring data and thus expedites the provisioning time. Experimental results show that our network-aware data transfer strategy offers the optimal solution when acquiring VMIs while introducing minimal overhead. Moreover, Nitro outperforms state-of-the-art VMI storage systems (eg, OpenStack Swift) by up to 77%.

In [22] we perform a performance evaluation of two communication bus mechanisms available in the openstack eco-system. Cloud computing depends on communication mechanisms implying location transparency. Transparency is tied to the cost of ensuring scalability and an acceptable request responses associated to the locality. Current implementations, as in the case of OpenStack, mostly follow a centralized paradigm but they lack the required service agility that can be obtained in decentralized approaches. In an edge scenario, the communicating entities of an application can be dispersed. In this context, we perform a study on the inter-process communication of OpenStack when its agents are geo-distributed. More precisely, we are interested in the different Remote Procedure Calls (OARPCs) implementations of OpenStack and their behaviours with regards to three classical communication patterns: anycast, unicast and multicast. We discuss how the communication middleware can align with the geo-distribution of the RPC agents regarding two key factors: scalability and locality. We reached up to ten thousands communicating agents, and results show that a router-based deployment offers a better trade-off between locality and load-balancing. Broker-based suffers from its centralized model which impact the achieved locality and scalability.

In [5], we give a complete overview of VMPlaceS, a dedicated framework we have been implementing since 2015 in order to evaluate and compare VM placement algorithms. Most current infrastructures for cloud computing leverage static and greedy policies for the placement of virtual machines. Such policies impede the optimal allocation of resources from the infrastructure provider viewpoint. Over the last decade, more dynamic and often more efficient policies based, e.g., on consolidation and load balancing techniques, have been developed. Due to the underlying complexity of cloud infrastructures, these policies are evaluated either using limited scale testbeds/in-vivo experiments or ad-hoc simulators. These validation methodologies are unsatisfactory for two important reasons: they (i) do not model precisely enough real production platforms (size, workload variations, failure, etc.) and (ii) do not enable the fair comparison of different approaches. More generally, new placement algorithms are thus continuously being proposed without actually identifying their benefits with respect to the state of the art. In this article, we present and discuss most of the features provided by VMPlaceS, a dedicated simulation framework that enables researchers (i) to study and compare VM placement algorithms from the infrastructure perspective, (ii) to detect possible limitations at large scale and (iii) to easily investigate different design choices. Built on top of the SimGrid simulation platform, VMPlaceS provides programming support to ease the implementation of placement algorithms and runtime support dedicated to load injection and execution trace analysis. To illustrate the relevance of VMPlaceS, we first discuss a few experiments that enabled us to study in details three well known VM placement strategies. Diving into details, we also identify several modifications that can significantly increase their performance in terms of reactivity. Second, we complete this overall presentation of VMPlaceS by focusing on the energy efficiency of the well-know FFD strategy. We believe that VMPlaceS will allow researchers to validate the benefits of new placement algorithms, thus accelerating placement research and favouring the transfer of results to IaaS production platforms. 

In [27], we present different heuristics that address the placement challenge in Fog/Edge infrastructures. As Fog Computing brings processing and storage resources to the edge of the network, there is an increasing need of automated placement (i.e., host selection) to deploy distributed applications. Such a placement must conform to applications’ resource requirements in a heterogeneous Fog infrastructure, and deal with the complexity brought by Internet of Things (IoT) applications tied to sensors and actuators. In this study, we present and evaluate four heuristics to address the problem of placing distributed IoT applications in the fog. By combining proposed heuristics, our approach is able to deal with large scale problems, and to efficiently make placement decisions fitting the objective: minimizing placed applications’ average response time. The proposed approach has been validated through comparative simulation of different heuristic combinations with varying sizes of infrastructures and applications.

In [35], we introduce the premises of monitoring function chaining conepts with the ultimate goal of delievering an holistic monitoring system for Fog/Edge infrastuctures. By relying on small sized and massively distributed infrastructures, the Edge computing paradigm aims at supporting the low latency and high bandwidth requirements of the next generation services that will leverage IoT devices (e.g., video cameras, sensors). To favor the advent of this paradigm, management services, similar to the ones that made the success of Cloud computing platforms, should be proposed. However, they should be designed in order to cope with the limited capabilities of the resources that are located at the edge. In that sense, they should mitigate as much as possible their footprint. Among the different management services that need to be revisited, we investigate in this study the monitoring one. Monitoring functions tend to become compute-, storage- and network-intensive, in particular because they will be used by a large part of applications that rely on real-time data. To reduce as much as possible the footprint of the whole monitoring service, we propose to mutualize identical processing functions among different tenants while ensuring their quality-of-service (QoS) expectations.We formalize our approach as a constraint satisfaction problem and show through micro-benchmarks its relevance to mitigate compute and network footprints.

In [17], we discuss the limitations of meta-data management in Fog/Edge infrastructures. A few storage systems have been proposed to store data in those infrastructures. Most of them are relying on a Distributed Hash Table (DHT) to store the location of objects which is not efficient because the node storing the location of the data may be placed far away from the object replicas. In this paper, we propose to replace the DHT by a tree-based approach mapping the physical topology. Servers look for the location of an object by requesting successively their ancestors in the tree. Location records are also relocated close to the object replicas not only to limit the network traffic when requesting an object, but also to avoid an overload of the root node. We also propose to modify the Dijkstra’s algorithm to compute the tree used. Finally, we evaluate our approach using the object store InterPlanetary FileSystem (IPFS) on Grid’5000 using both a micro experiment with a simple network topology and a macro experiment using the topology of the French National Research and Education Network (RENATER). We show that the time to locate an object in our approach is less than 15 ms on average which is around 20% better than using a DHT.

Cloud and HPC convergence

Geo-distribution of Cloud Infrastructures is not the only current trend of utility computing. Another important challenge is to favor the convergence of Cloud and HPC infrastructures, in other words on-demand HPC. Among challenges of this convergence is, for example, how to exploit HPC systems to execute data-intensive workflows effectively, as well as how to schedule tasks and jobs in Cloud, HPC, or hybrid HPC/Cloud infrastructures to meet data volatility and the ever-growing heterogeneity in the computation demands of workflows.

With the growing needs of users and size of data, commodity-based infrastructure will strain under the heavy weight of Big Data. On the other hand, HPC systems offer a rich set of opportunities for Big Data processing. As first steps toward Big Data processing on HPC systems, several research efforts have been devoted to understanding the performance of Big Data applications on these systems. Yet the HPC specific performance considerations have not been fully investigated. In [28], we conduct an experimental campaign to provide a clearer understanding of the performance of Spark, the de facto in-memory data processing framework, on HPC systems. We ran Spark using representative Big Data workloads on Grid’5000 testbed to evaluate how the latency, contention and file system’s configuration can influence the application performance. We discuss the implications of our findings and draw attention to new ways (e.g., burst buffers) to improve the performance of Spark on HPC systems.

Motivated by the our work [28] , we extend Eley  [107], a burst buffer solution that aims to accelerate the performance of Big Data applications, to be interference-aware. Specifically, while data prefetching reduce the response time of Big data applications as data inputs will be stored on a low-latency device close to computing nodes, it may come at a high cost for the HPC applications: the continuous interaction with the parallel file system (i.e., I/O read requests) may introduce a huge interference at the parallel file system level and thus end up with a degraded and unpredictable performance for HPC applications. In [7], we introduce interference and performance models for both HPC and Big Data applications in order to identify the performance gain and the interference cost of the prefetching technique of Eley; and demonstrate how Eley chooses the best action to optimize the prefetching while guaranteeing the pre-defined QoS requirement of HPC applications. For example, with 5% QoS requirement of the HPC application, Eley reduces the execution time of Big Data applications by up to 30% compared to the Naive burst buffer solution (NaiveBB) while guaranteeing the QoS requirement. On the other hand, the NaiveBB violates the QoS requirement by up to 58%.

Besides Clouds, Data Stream Processing (DSP) applications are widely deployed in HPC systems, especially the ones which require timely responses. DSP applications are often modelled as a directed acyclic graph: operators with data streams among them. Inter-operator communications can have a significant impact on the latency of DSP applications, accounting for 86% of the total latency. Despite their impact, there has been relatively little work on optimizing inter-operator communications, focusing on reducing inter-node traffic but not considering inter-process communication (IPC) inside a node, which often generates high latency due to the multiple memory-copy operations. In [26], we introduce a new DSP system designed specifically to address the high latency caused by inter-operator communications, called TurboStream. To achieve this goal, we introduce (1) an improved IPC framework with OSRBuffer, a DSP-oriented buffer, to reduce memory-copy operations and waiting time of each single message when transmitting messages between the operators inside one node, and (2) a coarse-grained scheduler that consolidates operator instances and assigns them to nodes to diminish the inter-node IPC traffic. Using a prototype implementation, we show that our improved IPC framework reduces the end-to-end latency of intra-node IPC by 45.64% to 99.30%. Moreover, TurboStream reduces the latency of DSP by 83.23% compared to JStorm.

Current data stream or operation stream paradigms cannot handle data burst efficiently, which probably results in noticeable performance degradation. In [25], we introduce a dual-paradigm stream processing, called DO (Data and Operation) that can adapt to stream data volatility. It enables data to be processed in micro-batches (ie, operation stream) when data burst occurs to achieve high throughput, while data is processed record by record (ie, data stream) in the remaining time to sustain low latency. DO embraces a method to detect data bursts, identify the main operations affected by the data burst and switch paradigms accordingly. Our insight behind DO’s design is that the trade-off between latency and throughput of stream processing frameworks can be dynamically achieved according to data communication among operations in a fine-grained manner (ie, operation level) instead of framework level. We implement a prototype stream processing framework that adopts DO. Our experimental results show that our framework with DO can achieve 5x speedup over operation stream under low data stream sizes, and outperforms data stream on throughput by 2.1 x to 3.2 x under data burst.

In the context of the Hydda project, where hybrid HPC/Cloud infrastructures are studied, heterogeneous dataflows, composed of coarse-grain tasks interconnected through data dependencies, are scheduled. Indeed, in heterogeneous dataflows, genomics dataflows for instance, some tasks may need HPC infrastructures (e.g., simulation) while other are suited for Cloud infrastructures (e.g., Big Data). Different quality of services are also expected from one task to the other. In [31] the scheduling of heterogeneous scientific dataflows is studied while minimizing the Cloud provider operational costs, by introducing a deadline-aware algorithm. Scheduling in a Cloud environment is a difficult optimization problem. Usually, works around the scheduling of scientific dataflows focus on public Clouds where the management of the infrastructure is an unknown black box. Thus, many works offer scheduling algorithms built to choose the best set of virtual machines through time such that the cost of the enduser is minimized. This paper presents a new algorithm based on HEFT that aims at minimizing the number of machines used by the Cloud provider, by taking deadlines into account.