Optimal application-oriented resource brokering in a high performance computing grid

Vincent Keller
EPFL, 2008
Thèse EPFL

Résumé

In the world of High Performance Computing (newly renamed "High Productivity Computing"), where the race for performance is on, where the hunt for the last Flop rages and where the developers swear by the Cult of Power, the different applications have different needs. Some of them are single threaded, others are parallel, some are memory bandwidth, others processor speed dominated. Some are embarrassingly parallel, point-to-point or multicast dominated, all depending on the algorithms and numerical methods they implement. Computing resources are also different, some are low-cost clusters of PlayStations 3, clusters with more or less efficient networks, others are expensive vector supercomputers. To express that diversity, a parameterization of the applications and the resources is proposed. Based on this parameterization, it is possible to quantify the needs of the different types of applications in terms of memory, processor and network requirements. On the other hand, the parameterization of the resources quantifies what a resource "offers" in term of memory, processor or network performance. Comparing the needs of the application and the offers of the resources, we present a model that predicts the computation and the communication times, and consequently the execution times of an application on different resources. We define a cost model that finds the best matches between an application and the available resources in an HPC Grid. It takes into account the cost of execution time of the application (based on the prediction previously mentioned), the waiting time cost (based on the information given by the local resource management systems), the cost of the licenses of the different software the application may use, the cost of the data transfer, and the ecological cost. The choice is made on a QOS (Quality of Service) given by the user, such as minimum turnaround time or minimum cost. Computing these different costs requires a fine knowledge of the behavior of the application on the resource. Thus we present a detailed application-oriented monitoring that consists in mapping information from the hardware monitoring and the different middleware the application (or the resource) uses, such as MPI detailed information, the hardware counters or the information of the execution on a local resource management system. The monitored data is stored in an archive system. The archived monitoring data concerning the behavior of the applications on the different resources, can then be used to help deciders purchasing new best suited resources for the current (and future) applications. In the second part of this work, we go further in matching the applications and resources: if it is possible to predict the behavior of an application on a given resource, it is also possible to fit the resource to the needs of the application. Knowing the processor performance needs of the application during its execution, it is possible to adjust its frequency through the frequency stepping mechanism that authorizes an operating system kernel or a program to modify on the fly the frequency (and the voltage) of the present days processors. Hence the resource exactly fits the needs of the application and the energy consumption of the processor is reduced. Finally, we propose a methodology to determine poorly implemented applications and thus giving users and developers hints and advice to detect flaws in an automatic manner. We present a real-life example, a spectral element code in CFD, for which the complexity changed when more than 1024 processors were used.

Official source

https://infoscience.epfl.ch/record/126278?ln=fr

À propos de ce résultat

Cette page est générée automatiquement et peut contenir des informations qui ne sont pas correctes, complètes, à jour ou pertinentes par rapport à votre recherche. Il en va de même pour toutes les autres pages de ce site. Veillez à vérifier les informations auprès des sources officielles de l'EPFL.

Concepts associés

Chargement

Publications associées

Chargement

Vincent Keller
EPFL, 2008
Thèse EPFL

Résumé

Official source

https://infoscience.epfl.ch/record/126278?ln=fr

À propos de ce résultat

Concepts associés (32)

Concepts associés

Chargement

Publications associées (66)

Publications associées

Chargement

Energy-Aware Processing Platform Exploration for Embedded Biosignal Analysis

Ahmed Yasir Dogan

According to the World Health Organization, lifestyle-related diseases, e.g., cardiovascular diseases are the major cause of mortality worldwide. An accurate and continuous medical supervision is highly required for diagnosis and treatment of such diseases. Our traditional healthcare delivery systems, however can’t cope with consequential increasing healthcare costs and medical management needs. Personal health monitoring systems are poised to offer large-scale and cost-effective solutions to this problem. The use of wearable, miniaturized and autonomous wireless sensor nodes, featuring continuous on-node analysis of biosignals, can indeed provide ambulatory long-term and real-time monitoring required by the patients, and enables faster coordination with medical personnel. In such autonomous nodes, due to very limited available energy resources and costly wireless transmission, an ultra-low-power (ULP) on-node processing platform for advanced biosignal analysis is crucial. In this thesis, I explore ULP processing architectures for on-node biosignal analysis applications; where commonly, moderately complex arithmetic manipulations on single- or multiple- input signals are carried out. To achieve energy efficiency while providing sufficient processing capability to apply advanced biosignal analysis, in this thesis near-threshold (near-Vt h ) computing is exploited. Hence, severe performance degradation and reliability issues, occurring at deeply scaled voltages, can be avoided. In Chapter 3, I introduce a near-Vth computing single-core architecture, consisting of a ULP core, an instruction memory (IM) and a data memory (DM). The ULP core features an instruction set architecture (ISA) customized for biosignal applications. I explore that an ISA with minimal instruction set achieves considerable energy savings compared to the state-of-the-art cores, when executing biosignal applications (i.e., up to 54% compared to an established ISA). The proposed single-core architecture accomplishes high energy efficiency for most of single-input biosignal analysis applications, since it fully exploits near-Vth computing. However, the single-core architecture achieves limited voltage scaling, hence reduced energy awareness, for most of multiple-input biosignal analysis applications, where computational workload requirements are such high that the single-core architecture can’t attain these throughputs in near-Vth regime. To alleviate the performance degradation issue that prevents the single-core architecture from exploiting near-Vt h computing typically for multiple-input biosignal analysis, I propose parallel processing of biosignals on multi-core architectures. To this end, In Chapter 4, a multiple instruction, multiple data (MIMD) multi-core architecture is introduced. The MIMD architecture comprises several ULP cores, individual IMs, and a multi-bank DM shared through a lightweight interconnect between the cores and the DM. I prove that parallel processing of multiple-input biosignals leads to better energy efficiency than the sequential processing (i.e., on a single-core) for moderate and high biosignal workloads. In particular, the MIMD architecture achieves up to 62% power savings with respect to the single-core architecture for high biosignal workloads (i.e., 167 MOps/s). On the other hand, parallel processing of multiple-input biosignals can be penalized at low workloads due to high leakage power dissipation in multi-core architectures. In particular, the MIMD architecture fails against the single-core architecture in terms of energy efficiency for workloads lighter than 1.7 MOps/s. One of the major burden of power dissipation in MIMD architectures is costly multiple instruction fetch. To mitigate this issue, I propose data-level parallelism through single instruction, multiple data (SIMD) paradigm. To this end, in Chapter 4 a novel hybrid multi-core architecture, that supports SIMD and MIMD operations, is introduced. The SIMD operations, coupled with data and instruction broadcasting, enable coordinated multiple accesses to memories, hence reduced instruction fetch power. Additionally, the hybrid multi-core architecture features partial power gating of memories to achieve leakage power savings, vital at low workloads (a few 100 kOps/s). I show that SIMD processing of multiple-input biosignals leads to better energy efficiency compared to the MIMD processing. In particular, when SIMD operations are exploited, the hybrid multi-core architecture achieves up to 45.7% power saving compared to the MIMD architecture for moderate biosignal workloads. I also ascertain that partial power gating of memories is an effective technique to alleviate leakage issue in multi-core architectures. More specifically, partial power gating of the IM in the hybrid multi-core architecture leads to 38.8% power saving at low workloads. Finally, to alleviate issues with applications involving such program parts that limit SIMD execution of applications (i.e., conditional program parts), I propose to resynchronize the cores for stable lockstep code execution in case of synchronization loss. Hence, SIMD operations are exploited even for applications with conditional program parts. To this end, in Chapter 4 a lightweight software-directed hardware synchronizer is introduced. I reveal that for applications with conditional program parts, lockstep SIMD execution accomplishes up to 64% power saving with respect to the elementary SIMD execution at moderate workloads (i.e.,89 MOps/s).

EPFL2013

Chargement

An Architecture for Load Balance in Computer Cluster Applications

Laurent Bindschaedler

Amid a data revolution that is transforming industries around the globe, computing systems have undergone a paradigm shift where many applications are scaled out to run on multiple computers in a computing cluster. As the storage and processing capabilities of a single machine are unable to keep pace with the amount of data, companies turn to distributed solutions to organize, persist, and analyze this Big Data. These new software solutions divide datasets into partitions that are processed in parallel on separate machines. A common problem in current cluster computing frameworks is load imbalance and limited parallelism due to skewed data distributions, processing times, and machine speeds. Load imbalance occurs when a few machines have more work and take longer to finish while the others remain idle, resulting in reduced overall performance and low resource utilization. The underlying cause for these issues is that data locality, where machines process the data stored locally, leads to tight coupling between a partition and the machine on which it is placed. This dissertation proposes a novel scatter architecture for computer cluster applications that effectively addresses the load imbalance problem and improves resource utilization. Scatter systems abandon data locality in favor of load balance. Existing systems first target locality, bringing computation to the data, and only then attempt to minimize load imbalance. Instead, we disregard any locality concerns and focus solely on achieving load balance by scattering all data across machines and bringing data to the computation as needed. The scatter architecture disaggregates compute and storage resources and pools all resources together across machines. We ensure storage load balance by spreading the data for each partition in small blocks across all storage devices and allow machines to retrieve data using an efficient, decentralized scheme. Scatter systems achieve compute load balance at runtime by dynamically adjusting the parallelism within a partition, either through work-sharing or by offloading background tasks to other machines. We design and implement three cluster applications inspired by the scatter architecture: a graph processing system that scales out using secondary storage, a general-purpose analytics framework, and a filesystem substrate that mitigates load imbalance for existing distributed databases. We demonstrate that each application can gracefully handle significant load imbalance with minimal performance degradation and that these applications can perform up to an order of magnitude faster than existing systems.

EPFL2020

Chargement

Complex networks exist for a number of purposes. The neural, metabolic and food networks ensure our survival, while the social, economic, transportation and communication networks allow us to prosper. Independently of the purposes and particularities of the physical embodiment of the networks, one of their fundamental functions is the delivery of information from one part of the network to another. Gossip and diseases diffuse in the social networks, electrochemical signals propagate in the neural networks and data packets travel in the Internet. Engineering networks for robust information flows is a challenging task. First, the mechanism through which the network forms and changes its topology needs to be defined. Second, within a given topology, the information must be routed to the appropriate recipients. Third, both the network formation and the routing mechanisms need to be robust against a wide spectrum of failures and adversaries. Fourth, the network formation, routing and failure recovery must operate under the resource constraints, either intrinsic or extrinsic to the network. Finally, the autonomously operating parts of the network must be incentivized to contribute their resources to facilitate the information flows. This thesis tackles the above challenges within the context of several types of networks: 1) peer-to-peer overlays – computers interconnected over the Internet to form an overlay in which participants provide various services to one another, 2) mobile ad-hoc networks – mobile nodes distributed in physical space communicating wirelessly with the goal of delivering data from one part of the network to another, 3) file-sharing networks – networks whose participants interconnect over the Internet to exchange files, 4) social networks – humans disseminating and consuming information through the network of social relationships. The thesis makes several contributions. Firstly, we propose a general algorithm, which given a set of nodes embedded in an arbitrary metric space, interconnects them into a network that efficiently routes information. We apply the algorithm to the peer-to-peer overlays and experimentally demonstrate its high performance, scalability as well as resilience to continuous peer arrivals and departures. We then shift our focus to the problem of the reliability of routing in the peer-to-peer overlays. Each overlay peer has limited resources and when they are exhausted this ultimately leads to delayed or lost overlay messages. All the solutions addressing this problem rely on message redundancy, which significantly increases the resource costs of fault-tolerance. We propose a bandwidth-efficient single-path Forward Feedback Protocol (FFP) for overlay message routing in which successfully delivered messages are followed by a feedback signal to reinforce the routing paths. Internet testbed evaluation shows that FFP uses 2-5 times less network bandwidth than the existing protocols relying on message redundancy, while achieving comparable fault-tolerance levels under a variety of failure scenarios. While the Forward Feedback Protocol is robust to message loss and delays, it is vulnerable to malicious message injection. We address this and other security problems by proposing Castor, a variant of FFP for mobile ad-hoc networks (MANETs). In Castor, we use the same general mechanism as in FFP; each time a message is routed, the routing path is either enforced or weakened by the feedback signal depending on whether the routing succeeded or not. However, unlike FFP, Castor employs cryptographic mechanisms for ensuring the integrity and authenticity of the messages. We compare Castor to four other MANET routing protocols. Despite Castor's simplicity, it achieves up to 40% higher packet delivery rates than the other protocols and recovers at least twice as fast as the other protocols in a wide range of attacks and failure scenarios. Both of our protocols, FFP and Castor, rely on simple signaling to improve the routing robustness in peer-to-peer and mobile ad-hoc networks. Given the success of the signaling mechanism in shaping the information flows in these two types of networks, we examine if signaling plays a similar crucial role in the on-line social networks. We characterize the propagation of URLs in the social network of Twitter. The data analysis uncovers several statistical regularities in the user activity, the social graph, the structure of the URL cascades as well as the communication and signaling dynamics. Based on these results, we propose a propagation model that accurately predicts which users are likely to mention which URLs. We outline a number of applications where the social network information flow modelling would be crucial: content ranking and filtering, viral marketing and spam detection. Finally, we consider the problem of freeriding in peer-to-peer file-sharing applications, when users can download data from others, but never reciprocate by uploading. To address the problem, we propose a variant of the BitTorrent system in which two peers are only allowed to connect if their owners know one another in the real world. When the users know which other users their BitTorrent client connects to, they are more likely to cooperate. The social network becomes the content distribution network and the freeriding problem is solved by leveraging the social norms and reciprocity to stabilize cooperation rather than relying on technological means. Our extensive simulation shows that the social network topology is an efficient and scalable content distribution medium, while at the same time provides robustness to freeriding.

EPFL2011

Chargement

EPFL2011

An Architecture for Load Balance in Computer Cluster Applications

Laurent Bindschaedler

EPFL2020

Energy-Aware Processing Platform Exploration for Embedded Biosignal Analysis

Ahmed Yasir Dogan

EPFL2013