Jaroslaw Slawinski

Unibus: Aspects of heterogeneity and fault tolerance in cloud computing

The paper describes our on-going project, termed Unibus, in the context of facilitating fault-tolerant executions of MPI applications on computing chunks in the cloud. In general, Unibus focuses on resource access virtualization and automatic, user-transparent resource provisioning that simplify use of heterogeneous resources available to users. In this work, we present the key Unibus concepts (the Capability Model, composite operations, mediators, soft and successive conditionings, metaapplications), and demonstrate how to employ Unibus to orchestrate resources provided by a commercial cloud provider into a fault-tolerant platform, capable of executing message passing applications. In order to support fault tolerance we use DMTCP [1] (Distributed MultiThreaded CheckPointing) that enables checkpointing at the users level. To demonstrate that the Unibus-created, FT-enabled platform allows to execute MPI applications we ran NAS Parallel Benchmarks and measured the overhead introduced by FT.

Enhancing Portability of HPC Applications across High-end Computing Platforms

Fast hardware turnover in supercomputing centers, stimulated by rapid technological progress, results in high heterogeneity among HPC platforms, and necessitates that applications are ported and adapted frequently. The cutting-edge nature of the hardware mandates customized performance tuning, which, coupled with continuously growing application complexity, makes the process inherently and increasingly challenging. In this paper, we analyze build procedures of a representative set of HPC applications, and attempt to identify commonalities that can be exploited to enhance cross-platform portability. We then propose a novel method for reducing non-portabilities while preserving high performance. The approach, based on profiles that capture and isolate non-portable features at various levels, requires only a moderate amount of changes to existing makefiles. It leverages the expertise of system designers and administrators, and reduces burdens placed on application scientists. As a proof of concept, we discuss the application of our methodology to enhancing portability of the Mile application across heterogeneous HPC platforms.

Platform and algorithm effects on computational fluid dynamics applications in life sciences

Abstract High Performance Computing (HPC) is a mainstream mode of exploration and analysis in different fields, not only technical but also social and life sciences. A well-established HPC domain is medicine, and cardiovascular sciences in particular. The adoption of CFD as a tool for diagnosis, prognosis, and treatment planning in the clinical routine is however still an open challenge. This computational tool, required by Computer Aided Clinical Trials and Surgical Planning, calls for significant computational resources to face both large volume of patients and diverse timelines ranging from election to emergency scenarios. Traditional local clusters may be not adequate to deliver the computational needs. Alternative solutions like grids and on-demand cloud resources need to be seriously considered. This paper proposes methodologies and protocols to identify the optimal choice of computing platforms for hemodynamics computations that will be increasingly needed in the future and the optimal scheduling of the tasks across the selected resources. We focus on hemodynamics in patient-specific settings and present extensive results on different platforms. We propose a way to measure and estimate performance and running time under realistic scenarios tailored to the utility function of the simulation. We discuss in detail the optimal (parallel) partitioning of the domain of a problem of interest with different mathematical approaches. We show that an overlapping splitting is generally advantageous and the detection of optimal overlapping has the potential to significantly reduce computational costs of the entire solution process and the communication volume across the platforms.

Issues in Communication Heterogeneity for Message-Passing Concurrent Computing

Heterogeneity in interconnection network throughput and latency has recently become a major issue for parallel computing applications. With the universal prevalence of multicore processors, large scale clusters and, most critically, cloud platforms, variations in communication characteristics of orders of magnitude are possible within a single execution environment. When applications also exhibit heterogeneity and irregularity in their communication patterns, process placement can make the difference between good and unacceptable performance. We discuss techniques for analyzing and addressing these factors in the context of a computational fluid dynamics application for the study of blood flow, on typical parallel platforms: a local parallel machine, a workstation network, and IaaS cloud-based cluster. Our experiences show problem sizes and platform sizes for which communication variations have significant impact, and suggest methods for process mapping that are likely to alleviate the detrimental effects of communication heterogeneity in different environments.

Extending Executability of Applications on Varied Target Platforms

High-performance applications are often developed for a specific class of target platforms and executing them on the increasing variety of Cloud and grid environments requires substantial adjustments and reconciliation. The aim of our framework, called ADAPT (Adaptive Application and Platform Translation), is to allow adaptation of applications for execution on various computational resources. The overall objective is to enhance usability of cyber-infrastructure platforms by providing automated adaptations of applications and conditioning of target environments so that greater cross-utilization is achieved. This paper presents a proof-of-concept experiment of a possible use of ADAPT, viz. executing a C/MPI application originally written for clusters on the Microsoft Azure infrastructure. This MPI application-to-cloud adjustment is based on automatic identification of the application programming paradigm followed by applying an application transformation to enable execution on an alternative target.

Portable builds of HPC applications on diverse target platforms

High-end machines at modern HPC centers are constantly undergoing hardware and system software upgrades - necessitating frequent rebuilds of application codes. The number of possible combinations of compilers, libraries, application build configurations, differing hardware architectures, etc, makes the process of building applications very onerous, requiring expert build knowledge from different domains. Our ongoing Harness Workbench Toolkit (HWT) project aims to foster and streamline the entire build process on heterogeneous computational platforms. This paper focuses on a key research issue of the HWT that regards facilitating and enhancement portability of build systems across multifarious machines, with particular respect to scientific software commonly used in the HPC community. The article presents a novel HWT approach based on the concept of generic build systems and profiles which encapsulate build knowledge provided independently by relevant experts. The paper describes profiles, the logistics of storing and retrieving build information, and interfacing to user-guided builds. We also report on experiences with applying the HWT approach to two scientific production codes (CPMD, GAMESS) on Cray XT4.

Unibus: a contrarian approach to grid computing

Abstract Despite maturing in many ways, heterogeneous distributed computing platforms continue to require substantial effort in terms of software installation and management for efficient use, often necessitating manual intervention by resource providers and end-users. In this paper we propose a novel model of resource sharing that is a viable alternative to that commonly adopted in the grid community. Our model, termed Unibus, shifts the resource virtualization and aggregation responsibilities to the software at the client side, taking these burdens away from resource providers. Drawing from parallels with operating systems, we argue that distributed resources may be unified and aggregated at the user’s end, in a manner similar to ordinary peripheral devices. Running on the user’s access device, the overlay system software can virtualize remote resources via dynamically deployed software mediators analogous to device drivers, reconfiguring the resources if necessary via “firmware” modules. To illustrate the feasibility of the Unibus model, we have prototyped a development toolkit automating the installation, build, run, and post-processing stages of MPI applications. Through the provided console, this toolkit can deploy and configure an MPI execution environment across a set of heterogeneous, isolated distributed resources, turning them into a coherent virtual machine with a single interface point. We conducted a series of experiments with the NAS Parallel Benchmarks. Results indicate that the toolkit preserves the application performance of “bare” MPI, while substantially reducing maintenance and configuration efforts. Overall, the results suggest that the envisioned client side overlay model for resource sharing may potentially be able to address some of long-standing obstacles in building heterogeneous HPC systems.

Adapting MPI to MapReduce PaaS Clouds: An Experiment in Cross-Paradigm Execution

One desired attribute of utility computing is the ability for any providers offering to meet any users requirement, but the variety of programming paradigms and platform models make this non-trivial. While higher specializations may be implemented on more generic layers, e.g. SaaS on PaaS, or PaaS on IaaS clouds, we attempt the inverse - deploying procedural message passing programs on a MapReduce platform. Although begun as an academic exercise, our experiences provide several insights into the feasibility of such a mapping and highlight some collateral benefits of deploying certain classes of MPI applications on MapReduce platforms. More generally, this potential for cross-paradigm execution marks a characteristic in the utility-like nature of cloud computing. Our approach is based on the concept of adapters, common in traditional utilities, to reconcile application requirements to platform facilities. Our design philosophy, middleware components, and results from a simple experiment are described.

Unibus-managed Execution of Scientific Applications on Aggregated Clouds

Our on-going project, Unibus, aims to facilitate provisioning and aggregation of multifaceted resources from resource providers and end-users’ perspectives. To achieve that, Unibus proposes (1) the Capability Model and mediators (resource drivers) to virtualize access to diverse resources, and (2) soft and successive conditioning to enable automatic and user-transparent resource provisioning. In this paper we examine the Unibus concepts and prototype in a real situation of aggregation of two commercial clouds and execution of benchmarks on aggregated resources. We also present and discuss benchmarks’ results.

Experiences with Cost and Utility Trade-offs on IaaS Clouds, Grids, and On-Premise Resources

Cloud computing is now a mainstream technology in many application domains and user constituencies. For scientific high-performance applications in academic and research settings however, the trade-offs between cost and elasticity on the one hand, and performance and access on the other, is not always clear. We discuss our experiences with comparing cost and utility for a hemodynamics computational fluid dynamics code on three typical platform-types available to researchers: IaaS clouds, grids, and on-premise local resources. To rank the tested platforms, we introduce a simple utility function describing the value of a completed computational task to the user as a function of the wait time and the cost of the computation. Our results suggest that IaaS clouds can be a convenient choice for the considered class of CFD simulations, providing a valuable trade-off between cost and task completion time.