Christos K. Filelis-Papadopoulos

CLOUDLIGHTNING: A Framework for a Self-organising and Self-managing Heterogeneous Cloud

As clouds increase in size and as machines of different types are added to the infrastructure in order to maximize performance and power efficiency, heterogeneous clouds are being created. However, exploiting different architectures poses significant challenges. To efficiently access heterogeneous resources and, at the same time, to exploit these resources to reduce application development effort, to make optimisations easier and to simplify service deployment, requires a re-evaluation of our approach to service delivery. We propose a novel cloud management and delivery architecture based on the principles of self-organisation and self-management that shifts the deployment and optimisation effort from the consumer to the software stack running on the cloud infrastructure. Our goal is to address inefficient use of resources and consequently to deliver savings to the cloud provider and consumer in terms of reduced power consumption and improved service delivery, with hyperscale systems particularly in mind. The framework is general but also endeavours to enable cloud services for high performance computing. Infrastructure-as-a-Service provision is the primary use case, however, we posit that genomics, oil and gas exploration, and ray tracing are three downstream use cases that will benefit from the proposed architecture.

Solving finite difference linear systems on GPUs: CUDA based Parallel Explicit Preconditioned Biconjugate Conjugate Gradient type Methods

During the last decades, explicit approximate inverse preconditioning methods have been used for efficiently solving sparse linear systems on multiprocessor systems. The effectiveness of explicit approximate inverse preconditioning schemes relies on the use of efficient preconditioners that are close approximants to the coefficient matrix and are fast to compute in parallel. A new parallel computational technique is proposed for the parallelization of the explicit preconditioned conjugate gradient type method on a Graphics Processing Unit (GPU). The proposed parallel methods have been implemented using Compute Unified Device Architecture (CUDA) developed by NVIDIA. The inherently parallel operations between vectors and matrices involved in the explicit preconditioned biconjugate conjugate gradient type schemes exhibit significant amounts of loop-level parallelism because of the matrix–vector and the vector–vector products that can lead to high performance gain on the GPU systems, specifically designed for such computations. Finally, numerical results for the performance of the explicit preconditioned biconjugate conjugate gradient type method for solving characteristic two dimensional boundary value problems, using the finite difference method, on a massive multiprocessor interface on a GPU are presented. The CUDA implementation issues of the proposed method are also discussed.

Managing Spatial Graph Dependencies in Large Volumes of Traffic Data for Travel-Time Prediction

The exploration of the potential correlations of traffic conditions between roads in large urban networks, which is of profound importance for achieving accurate traffic prediction, often implies high computational complexity due to the implicated network topology. Hence, focal methods are required for dealing with the urban network complexity, reducing the performance requirements that are associated to the classical network search techniques (e.g., Breadth First Search). This paper introduces a graph-theory-based technique for managing spatial dependence between roads of the same network. In particular, after representing the traffic network as a graph, the local neighbors of each road are extracted using Breadth First Search graph traversal algorithm and a lower complexity variant of it. A Pearson product-moment correlation-coefficient-based metric is applied on the selected graph nodes for a prescribed number of level sets of neighbors. In order to evaluate the impact of the new method to the traffic prediction accuracy achieved, the most correlated roads are used to build a STARIMA model, taking also into account the possible time delays of traffic conditions between the interrelated roads. The proposed technique is benchmarked using traffic data from two different cities: Berlin, Germany, and Thessaloniki, Greece. Benchmark results not only indicate significant improvement on the computational time required for calculating traffic correlation metric values but also reveal that a different variant works better in different network topologies, after comparison to third-party approaches.

GENERIC APPROXIMATE SPARSE INVERSE MATRIX TECHNIQUES

During the last decades explicit preconditioning methods have gained interest among the scientific community, due to their efficiency for solving large sparse linear systems in conjunction with Krylov subspace iterative methods. The effectiveness of explicit preconditioning schemes relies on the fact that they are close approximants to the inverse of the coefficient matrix. Herewith, we propose a Generic Approximate Sparse Inverse (GenASPI) matrix algorithm based on ILU(0) factorization. The proposed scheme applies to matrices of any structure or sparsity pattern unlike the previous dedicated implementations. The new scheme is based on the Generic Approximate Banded Inverse (GenAbI), which is a banded approximate inverse used in conjunction with Conjugate Gradient type methods for the solution of large sparse linear systems. The proposed GenASPI matrix algorithm, is based on Approximate Inverse Sparsity patterns, derived from powers of sparsified matrices and is computed with a modified procedure based on the GenAbI algorithm. Finally, applicability and implementation issues are discussed and numerical results along with comparative results are presented.

A class of generic factored and multi-level recursive approximate inverse techniques for solving general sparse systems

Purpose – The purpose of this paper is to propose novel factored approximate sparse inverse schemes and multi-level methods for the solution of large sparse linear systems. Design/methodology/approach – The main motive for the derivation of the various generic preconditioning schemes lies to the efficiency and effectiveness of factored preconditioning schemes in conjunction with Krylov subspace iterative methods as well as multi-level techniques for solving various model problems. Factored approximate inverses, namely, Generic Factored Approximate Sparse Inverse, require less fill-in and are computed faster due to the reduced number of nonzero elements. A modified column wise approach, namely, Modified Generic Factored Approximate Sparse Inverse, is also proposed to further enhance performance. The multi-level approximate inverse scheme, namely, Multi-level Algebraic Recursive Generic Approximate Inverse Solver, utilizes a multi-level hierarchy formed using Block Independent Set reordering scheme and an a...

A Generic Framework Supporting Self-Organisation and Self-Management in Hierarchical Systems

A novel, generic, framework for supporting self-organisation and self-management in hierarchical systems is presented. The framework allows for the incorporation of local self-organising and self-managing strategies at each level in the hierarchy. These local strategies determine the behaviour of that level and the effects of these strategies can be communicated to, and used by, the strategies in adjacent levels of the hierarchy. Thus, in general, strategies may be viewed as parameterised functions. Information emanating from both the lower and the upper levels in the hierarchy can be used as parameters. Information from below represents the status of the lower levels, whereas information from above can be used to influence the direction and the rate of system evolution. As the component parts of the system evolve to their goal states, the rate of evolution slows. At that point, by definition, a component is maximally contributing to the global goal state of the system as a whole. A novel concept to measure the distance that a component is from this stasis, its Suitability Index is presented and formally defined. Although the proposed framework can be generalised to any hierarchical system, this paper applies it specifically to large scale, hierarchically structured, computer systems. An implementation of this framework and an empirical study of its effectiveness has been conducted as part of the the CloudLightning Project.

Large-scale simulation of a self-organizing self-management cloud computing framework

A recently introduced cloud simulation framework is extended to support self-organizing and self-management local strategies in the cloud resource hierarchy. This dynamic hardware resource allocation system is evolving toward the goals defined by local strategies, which are determined as maximization of: energy efficiency of cloud infrastructures, task throughput, computational efficiency and resource management efficiency. Heterogeneous hardware resources are considered that are except from commodity CPU servers, hardware accelerators such as GPUs, MICs and FPGAs, thus forming a heterogeneous cloud infrastructure. Energy consumption and task execution models for the heterogeneous accelerators are also proposed, in order to demonstrate the energy efficiency of the proposed resource allocation system. Implementation details of the new functionalities on the parallel cloud simulation framework are discussed, while numerical results are given for the scalability and utilization of the cloud elements using the self-organization and self-management framework with two VM placement strategies.

A framework for simulating large scale cloud infrastructures

Abstract Cloud infrastructures are continuously growing in size, since more cloud nodes are added to already existing hyper-scale infrastructures. These hyper-scale infrastructures are also becoming heterogeneous as different types of accelerators are added in order to increase performance per watt for certain types of applications and allow for various HPC workloads to migrate to Cloud environments. The introduction of diverse workloads that migrate in the Cloud along with increasing volume of incoming tasks results in phenomena of network congestion, underutilization and resource fragmentation. Simulators are used to analyze, study and possibly improve Cloud environments. However, existing Cloud simulation tools lack the ability to handle heterogeneous resources and tasks that span across multiple Cloud nodes. Moreover, they are mostly sequential and cannot scale to large numbers of Cloud nodes. Furthermore, they do not support over-commitment, which is a common practice in real-world Cloud environments. A framework for simulating large numbers of heterogeneous cloud nodes organized in Cells and executing large numbers of HPC tasks is proposed. The framework is inherently parallel and designed for hybrid distributed memory parallel systems, supporting CPU, memory and network over-commitment. The simulation framework is based on a time advancing loop, allowing dynamic change of the granularity of the simulator and minimizing memory requirements, since data related to the current time-step is stored. Moreover, a latency model for the currency of data in the Gateway Service and Broker is also supported. Implementation details along with discussions concerning the extensibility of the framework are given. Numerical results for simulating large number of heterogeneous resources and incoming tasks are also presented.

On the power consumption modeling for the simulation of Heterogeneous HPC clouds

During the last years, except from the traditional CPU based hardware servers, hardware accelerators are widely used in various HPC application areas. More specifically, Graphics Processing Units (GPUs), Many Integrated Cores (MICs) and Field-Programmable Gate Arrays (FPGAs) have shown a great potential in HPC and have been widely mobilized in supercomputing. With the adoption of HPC from cloud environments, the realization of HPC-Clouds is evolving since many vendors provide HPC capabilities on their clouds. With the increase of the interest on clouds, there has been an analogous increase in cloud simulation frameworks. Cloud simulation frameworks offer a controllable environment for experimentation with various workloads and scenarios, while they provide several metrics such as server utilization and power consumption. For providing these metrics, cloud simulators propose mathematical models that estimate the behavior of the underlying hardware infrastructure. This paper focuses on the power consumption modeling of the main compute elements of heterogeneous HPC servers, i.e. CPU servers and pairs of CPU-accelerators. The modeling approaches of existing cloud simulators are examined and extended, while new models are proposed for estimating the power consumption of accelerators.

On numerical modeling performance of generalized preconditioned methods

During the last decades, research efforts have been focused on the derivation of effective preconditioned iterative methods. The preconditioned iterative methods are mainly categorized into implicit preconditioned methods and explicit preconditioned methods. In this manuscript we review implicit preconditioned methods, based on incomplete and approximate factorization, and explicit preconditioned methods, based on sparse approximate inverses and explicit approximate inverses. Modified Moore-Penrose conditions are presented and theoretical estimates for the sensitivity of the explicit approximate inverse matrix of the explicit preconditioned method are derived. Finally, the performance of the preconditioned iterative methods is illustrated by solving characteristic 2D elliptic problem and numerical results are given indicating a qualitative agreement with the theoretical estimates.