joint-lab workshop Jun. 9-11 2014

Wolfgang Frings

When trying to make auto-tuning practical using common infrastructure, public repository of knowledge, and machine learning (cTuning.org), we faced a major problem with reproducibility of experimental results collected from multiple users. This was largely due to a lack of information about all software and hardware dependencies as well as a large variation of measured characteristics.
I will present a possible collaborative approach to solve above problems using a new Collective Mind knowledge management system. This modular infrastructure is intended to preserve and share through Internet the whole experimental setups with all related artifacts and their software and hardware dependencies besides just performance data. Researchers can take advantage of shared components and data with extensible meta-description at http://c-mind.org/repo to quickly prototype and validate research techniques particularly on software and hardware optimization and co-design. At the same time, behavior anomalies or model mispredictions can be exposed in a reproducible way to interdisciplinary community for further analysis and improvement. This approach supports our new open publication model in computer engineering where all results and artifacts are continuously shared and validated by the community (c-mind.org/events/trust2014).

Xin Zhao

Programming Runtime Support for Irregular Computations

Irregular computations have become increasingly important in many areas in recent year such as bioinformatics and social network analysis. Traditional data movement approaches for scientific computation are not well suited for such applications. The Active Messages (AM) model is an alternative communication paradigm that is better suited for such applications by allowing computation to be dynamically moved closer to data. Given the wide usage of MPI in scientific computing, enabling an MPI-interoperable AM paradigm would allow traditional applications to incrementally start utilizing AMs in portions of their applications, thus eliminating the programming effort of rewriting entire applications.

In our previous work we proposed a new generalized framework for MPI-interoperable Active Messages that can provide rich semantics to accommodate a wide variety of application computational patterns. Together with a new API, we present a detailed design of the correctness semantics of the functionality, including memory semantics, interoperability, ordering, concurrency, etc. We also proposed techniques for data streaming, buffering management and asynchronous processing to guarantee the correct execution of irregular applications as well as to achieve high performance. In this talk, I will discuss about irregular computations and the effort we made from programming model and runtime to make the computations easier and faster.

Xin Zhao is a fourth-year Ph.D. student from the Department of Computer Science at the University of Illinois at Urbana-Champaign (UIUC), advised by Prof. William Gropp. Her research interests focus on parallel programming models / runtime systems and irregular applications, with an emphasis on communication, resources management and dynamic execution.

Victor Lopez

DLB: Dynamic Load Balancing Library

Distribute equal amounts of work between tasks is not always trivial and usually becomes a negative performance impact in an application. DLB is a dynamic library designed to speed up hybrid applications by improving its load balance with little or none intervention from the user. The idea behind the library is to redistribute the computational resources of the second level of parallelism (OpenMP, OmpSs) to improve the load balance of the outer level of parallelism (MPI). DLB library uses an interposition technique at run time, so it is not necessary to do a previous analysis or modify the application; although finer control is also supported through an API.

We will present also a case study with CESM (Community Earth System Model), a global climate model that provides computer simulations of the Earth climate states. The application already uses a hybrid parallel programming model (MPI+OpenMp), so with few modifications in the source code we have compiled it to use the OmpSs programming model where DLB will benefit from the high malleability of it.

Marc Snir

Runtime and OS research at DoE

Pieter Bellens

Quantifying the effect of rectangular blocks in the dense QR factorization

Blocked, dense QR factorization using Householder reflectors attains minimal communication bounds and creates a fine-grained parallel computation. We consider the effects of rectangular block dimensions for an implementation in OmpSs, hereby unifying the traditional algorithm, that uses panels or block columns, and the square-blocked variants. Communication, computation, potential parallelism and hence the performance are functions of the block dimension. We use hardware counters and the Task Dependence Graph to quantize these properties for different matrix dimensions. Our measurements indicate that, against the grain of traditional practice, performance in dynamically scheduled environments can be improved by resorting to blocks with rectangular dimensions

Francois Tessier

Distributed communication-aware load balancing with TreeMatch in Charm++

Programming multicore or manycore architectures is a hard challenge particularly if one wants to fully take advantage of their computing power. Moreover, a hierarchical topology implies that communication performance is heterogeneous and this characteristic should also be exploited. We developed a parallel and distributed hierarchical load balancer for Charm++ that take into account both aspects. This work is based on our TreeMatch library that computes process placement in order to reduce an application communication cost based on the hardware topology. We show that the proposed load-balancing scheme manages to improve the execution times while being computed fast and in a scalable manner.

Brice Videau

Porting HPC applications to the Mont-Blanc prototype using BOAST

One of the goal of the Mont-Blanc project is to use real HPC application to evaluate the feasibility of exascale architectures using off the shelf hardware commonly used in the embedded world. The porting of those application is thus of paramount importance for the project. But, if getting scientific applications to run on the target platform is not very difficult, obtaining good performance portability is challenging. Indeed, HPC software are often hand tuned for the most frequently encountered architectures, and those optimizations can prove harmful if applied on a very different architecture. One way to alleviate this problem is to use task based runtimes to obtain adaptive application from a load balancing and network point of view. Unfortunately this solves only part of the problem. Individual tasks also have to be optimized and can be very sensitive to many parameters that are often not clearly exposed in the source code. We thus propose BOAST a meta-programming tool aiming at generating parametrized source code. Several output languages are supported and an expressive DSL is defined to help express the optimizations. An integrated compilation and execution framework is also supplied. This allows to directly test the generated kernels inside BOAST. This talk will present BOAST and how we used it to port part of two HPC applications:

- the Debauchies wavelet kernels of BigDFT, a quantum physics software that compute the electronic density around atoms and molecules,

- a port from CUDA to OpenCL of SPECFEM3D_GLOBE, a wave propagation software based on spectral finite element  methods.

Performance results will also be presented.

Lokman Rahmani

Smart In Situ Visualization for Climate Simulations 

 The increasing gap between computational power and I/O performance in new supercomputers has started to drive a shift from an offline approach to data analysis to an inline approach, termed in situ visualization (ISV). While most visualization software now provides ISV, they typically visualize large dumps of unstructured data, by rendering everything at the highest possible resolution. This often negatively impacts the performance of simulations that support ISV, in particular when ISV is performed interactively, as in situ visualization requires synchronization with the simulation. In this work, we advocate for a smarter method of performing ISV. Our approach is data-driven: it aims to detect potentially interesting regions in the generated dataset in order to feed ISV frameworks with “the interesting” subset of the data produced by the simulation. While this method mitigates the load on ISV frameworks by making them more efficient and more interactive, it also helps scientists focus on the relevant part of their data. We investigate smart ISV in the context of a climate simulation, with a set of generic filters derived from information theory, statistics and image processing, and show the tradeoff between performance and quality of visualization.

Justin Wozniak

Case Studies in Big Data and HPC from X-ray Crystallography

Recent advancements in X-ray crystallography methods, including experimental techniques and detector technology, have produced a data explosion (10's of TBs/week) that has outpaced increases in conventional computational and storage capacity, leading to a crisis in computational analysis and data management in X-ray sciences. Typical Big Data solutions do not accommodate the ad hoc nature of the scientific workflow, including opportunistic use of hardware and highly specialized analysis tools.  From a computational perspective, existing analysis codes must be quickly scaled up to massively parallel resources.  In this presentation, we will describe our recent work applying the Swift programming language to four applications in X-ray sciences, addressing problems in wide-area data movement and management as well as scaling existing applications on large clusters and the Blue Gene/Q.

Christine Morin

Contrail: Interoperability and Dependability in a Cloud Federation

Cloud computing market is in rapid expansion due to the opportunities to dynamically allocate large amount of resources when needed and to pay only for their effective usage. However, many challenges, in terms of interoperability, performance guarantee, and dependability, should still be addressed to make cloud computing the right solution for companies. Contrail integrated project (IP), funded by the European Commission (http://www.contrail-project.eu) developed a comprehensive cloud computing software stack in open source to address these challenges.
In this talk we first discuss the main challenges faced in the open cloud market and then we present components developed in the framework of the Contrail European project to provide solutions to guarantee interoperability in a cloud federation and to deploy distributed applications over a federation of heterogeneous cloud providers. Our solutions allow to negotiate QoS and QoP SLA terms for an application and then map them on the physical resources. 

Martin Quinson

Formal verification of unmodified MPI applications with SimGrid

This talk will first recap the approach leveraged in SimGrid to formally assess the correction of MPI applications through model checking. It will be focused on our current status report and future work. We are now able to verify safety properties, but also liveness properties (with some restrictions), on unmodified small to medium MPI applications (few thousands of lines in C, C++ or Fortran). I will conclude with the research leads that we are currently working on, and with the kind of collaboration that could occur within the Joint Lab with the potential users of such tools.

Yves Robert

Algorithms for coping with silent errors
Silent errors have become a major problem for large-scale distributed systems. Detection is hard, and correction is even harder. This talks presents generic algorithms to achieve both detection and correction of silent errors, by coupling verification mechanisms and checkpointing protocols.

Slim Bouguerra

Energy-Performance Tradeoffs in Multilevel Checkpoint Strategies

 Increased complexity of computer architectures, consideration of power constraints, and expected failure rates of hardware components make the design and analysis of energy-efficient fault-tolerance schemes an increasingly challenging and important task. We develop run-time and energy models for multilevel checkpoint schemes and characterize when tradeoffs between expected runtime and energy usage exist. Using these models, we study FTI, a recently developed multilevel checkpoint library, on an IBM Blue Gene/Q. We show that FTI has a low energy footprint and that, consequently optimal checkpoint-interval values with respect to time and energy are similar. We also explore the effect of general system-level parameters on run-time and energy tradeoffs.

Tatiana V. Martsinkevich

Using dedicated resources to alleviate memory limitation for message logging protocols

There are different approaches on how to handle memory limitation for a message logging protocol. The simplest is to take a checkpoint once one of the processes runs out of memory or dump logs to the stable storage to free the memory. However this may increase the load on the the I/O subsystem which is not desirable especially for large-scale runs . Another approach is to use the memory of additional dedicated nodes as a log storage: when a process runs out of memory it sends a portion of its log to the memory of a dedicated node. I will present the study on the feasibility of this approach and explore the overheads related to it.

Ana Gainaru

The road to failure prediction on Blue Waters: latest details and future directions

We analyze the characteristics of failures from the Blue Waters system and study their effect on the results given by the online failure prediction. We make a couple of key observations about the difference in behaviour between different failure types and propose specific optimizations for each. A detailed analysis of the prediction results is also given. We present future work direction together with preliminary results.

Leonardo Bautista Gomez

Fault Tolerance Interface new features and new developments

Slim Bouguerra

Energy-Performance Tradeoffs in Multilevel Checkpoint Strategies

Increased complexity of computer architectures, consideration of power constraints, and expected failure rates of hardware components make the design and analysis of energy-efficient fault-tolerance schemes an increasingly challenging and important task. We develop run-time and energy models for multilevel checkpoint schemes and characterize when tradeoffs between expected runtime and energy usage exist. Using these models, we study FTI, a recently developed multilevel checkpoint library, on an IBM Blue Gene/Q. We show that FTI has a low energy footprint and that, consequently optimal checkpoint-interval values with respect to time and energy are similar. We also explore the effect of general system-level parameters on run-time and energy tradeoffs.