SuperB Computing R&D Workshop

9 Mar 2010, 14:00 → 12 Mar 2010, 18:00 Europe/Rome

For details, see the Workshop Web site at http://www.fe.infn.it/superb

Tuesday, 9 March

14:30 → 16:30 Plenary Session: Introduction

14:30 Goals of the workshop

Speaker: Mauro Morandin (INFN)

Slides Morandin_introduction.pdf

14:50 The BaBar computing model and what we are likely to preserve in SuperB

Speaker: Dr David Brown (Lawrence Berkeley National Lab)

abstract babar_abstract.txt

Slides BaBar_Reuse.key BaBar_Reuse.pdf

15:20 SuperB and its computing requirements

Speaker: Fabrizio Bianchi (TO)

Slides BianchiSuperBCompRequirements.pdf BianchiSuperBCompRequirements.pptx

15:50 Discussion

16:30 → 17:00 Coffee Break

17:00 → 19:00 Plenary Session: Impact of new CPU architectures (I)

Exploitation of modern CPU architectures and its impact on the computing model of HEP experiments

17:00 Introduction and goals

Speaker: Vincenzo Innocente (CERN)

pdf SuperBIntro.pdf

Slides SuperBIntro.pptx

17:15 Modern CPU architectures: Challenges and opportunities for HEP

In the last few years Computing has been caracterized by the advent of "multicore cpus". Effective exploitation of this new kind of computing architecture requires the adaptation of legacy software and enventually a shift of the programming paradigms to massive parallel. In this talk we will introduce the reasons that brough to the introduction of "multicore" hardware and the consequencies on system and application software. The activities initiated in HEP to adapt current software will be reviewed before presenting the perspective for the future.

Speaker: Vincenzo Innocente (CERN)

pdf VIatSuperB0310.pdf

Slides VIatSuperB0310.ppt

18:15 Experience and perspective in parallelization of data analysis (including experience from BaBar and future plan)

Increasing the dimension of data samples for data analyses and using advanced algorithms for background suppression (such as unbinned maximum likelihood fits) require high CPU performance. Recently, vendors like Intel and AMD have not incremented the performance of single CPU unit as in the past, but they are working on multi-core CPU. Currently we have up to 8 cores implemented on one single chip. Other possibilities are offered by the use of Graphical Process Units (GPUs), which offer hundreds of cores. This fact represents a possible revolution in the development of new programs. Indeed we can parallelize the code obtaining great benefits from new multi-core architectures. So we have to reformulate some algorithms generally used for data analyses in High Energy Physics (e. g. Minuit for maximum likelihood fits). It is also possible to have faster execution of the code using the Message Passing Interface (MPI) paradigm, spreading the execution of the code over several CPUs in a cluster. These techniques of High Performance Computing (HPC) are well established in other fields, like computational chemistry and astrophysics. In High Energy Physics community there is not such a large use, but in the future it can be an elegant solution in all the cases where the data analyses will get more and more complicated. In particular the huge data samples expected to be collected in the next generations of experiments, such as the SuperB-factories, will represent a real challenge for the data analysis, requiring to rewrite the algorithms in parallel. In this presentation I will give a description of the data analysis softwares used in the BaBar experiment and, taking experience from their complexity, I will give an overview on the possible scenario for the next generation of SuperB factories.

Speaker: Alfio Lazzaro (MI)

Slides lazzaro.pdf lazzaro.pptx

18:45 Discussion

Wednesday, 10 March

08:30 → 10:30 Plenary Session: Impact of new CPU architectures (II)

08:30 Coding for the GPU: highlights and application examples

Currently, technological solutions being adopted by more manufacturers are bringing to CPU architecture with an even more degree of parallelism, evolving from the multi-core era to the "many-core" era. In this scenario hundreds and, in short, thousands of processing cores are contained within the same processor. A so deep change in architectural paradigm compels an equally deep change in algorithms and programming paradigms. In this context, the GPU (Graphics Processing Unit) were created to accelerate typical 2D and 3D graphic processing, characterized by an extreme degree of parallelism, containing hundreds of processing cores. Due to their characteristics, these are now used to perform complex calculations also in more general fields. During the talk an introduction to GPU Computing will be presented, providing a small technological outlook on manycore scenarios and a brief introduction on architectures and programming languages such as CUDA and OpenCL. We report performance information obtained with algorithms developed in Perugia in the context of the European project Einstein Telescope (ET) and the INFN project macgate (Manycore Computing for Future Gravitational Observatories) and with others of more common interest, such as Monte Carlo methods, Random Number Generation and Simulations.

Speaker: Leone Battista Bosi (PG)

Slides 2010-Mar-9-Ferrara_BOSI.pdf

09:00 Exploiting concurrency to speed-up ROOT data analysis: caching, PROOF and multi-threading

Concurrency aims to improve computing performance by executing a set of computations simultaneously, possibly in parallel. Since the advent of today's many-core machines the full exploitation of the available CPU power has been one of the main challenges for high-performance computing software projects, including the HEP ones. However, in HEP data analysis the bottleneck is not (only) CPU but also the I/O hardware. In this talk I will discuss what is done in ROOT improve the I/O performance exploiting the technological advancements. This starts by optimizing the way information is stored in files to improve the readout performance using cache techniques. Then a multi-process approach √ la PROOF allows to exploit the full I/O bandwidth on multi-core machines; more generally, the PROOF technology can effectively increase the I/O capabilities of LAN (and perhaps WAN) setups. Finally the use of multiple threads may help in reducing the time needed to get the data available in the application: I will present the current status of the parallel unzipping implementation, and discuss possible improvements and applications to the write case.

Speaker: Gerry Ganis

Slides Ganis-Ferrara.pdf

09:30 Multi-level Parallel fit algorithms using MPI and CUDA

Linux clusters consisting of multi-core commodity-chip based nodes, augmented with GPGPU accelerators are becoming common at computing centers. In this talk we report on some early results of a study to investigate use of this computing paradigm to accelerate the fitting algorithms used in MINUIT. In particular we show that very good speedups are possible for the negative log likelihood (NLL) fit of a simple Guassian. We also discuss a preliminary implementation of Breit-Wigner convoluted with a Guassian and report on the challenges encountered implementing the complex error function of this algorithm on the GPU.

Speaker: Karen Tomko

Slides SuberB_Tomkorev3.pdf

09:50 Discussion

10:30 → 11:00 Coffee Break

11:00 → 13:00 Plenary Session: Software architecture and framework

11:00 Introduction and goals

Speaker: Peter Elmer (Princeton University)

11:15 Framework evolution: status and outlook

Object-Oriented event data processing software frameworks have been developed by all recent HEP experiments to structure and support their data processing programs such as the simulation, reconstruction and data analysis.  These frameworks implement a given architectural vision and provide a set of features and functionalities that are very close among the various experiments. This presentation will try to identify which are these common architectural elements and the functionalities of an ideal framework based on the experience of the frameworks in use at the LHC experiments.

Speaker: Dr Pere Mato (CERN)

Slides FrameworksEvolution-20100310.pptx

11:45 Discussion

13:00 → 14:30 Lunch

14:30 → 16:30 Plenary Session: Persistence and data handling models

14:30 Introduction and goals

Speaker: Dr David Brown (Lawrence Berkeley National Lab)

abstract data_abstract.rtf

Slides Data_intro.pdf

14:45 Persistency in HEP applications: current models and future outlook

Speaker: Paolo Calafiura (Lawrence Berkeley National Lab)

Slides SuperB_Persistency_Mar10.pdf

15:15 The SciDB approach and its applicability to HEP computing

SciDB is a new open source database management system that emerged from the Extremely Large Database (XLDB) workshop series.  Specifically designed to meet the needs of many scientific disciplines, it features an array-based data model that adds order and adjacency to the traditional relational set model.  SciDB's array model and the SciDB software in its currently-projected form is *not* appropriate for storage of the largest HEP data type (event data), as I understand it. So why discuss it at this workshop?  Aside from its potential uses for more structured, though smaller, data sets, I believe there are several lessons to be learned from the design and implementation of SciDB that are generally applicable to large-scale scientific data management and analysis.

Speaker: K.-T. Lim

Slides SuperB-KTL.pdf

15:55 Discussion

16:30 → 17:00 Coffee Break

17:00 → 19:15 Plenary Session: Code development: languages, tools, standards, QA, interplay with Online

17:00 Introduction and goals

Speakers: Andrea Di Simone (RM2), Prof. Roberto Stroili (Universita' di Padova and INFN), Dr Steffen Luitz (SLAC)

Slides luitz-intro.pdf stroili-intro.pdf

17:20 Release building and validation at LHC

Presentation will be summary of release building and validation procedures done in LHC experiments at CERN. Brief description how this is done in LHC experiments with main focus on ATLAS. The complex detectors determine the existing large collaboration with many users and developers providing software code. Every experiment has its own developed and used tools but there are also many common utilities and code used and shared between the experiments.ATLAS software comprises around 2000 software packages, organized in 10 projects that are build on a variety of compiler and operating system combinations every night. File-level parallelism, package-level parallelism and multi-core build servers are used to perform simultaneous builds of 6 platforms that are merged into single installation. I will discuss the various tools that provide performance gains and the error detection and retry mechanisms that have been developed in order to counteract network and other instabilities. Will be discussed test tools and the different levels of testing and validation framework which are implemented in order to give better quality view of the produced software.

Speaker: Emil Obreshkov (CERN)

Slides ReleaseB&V@LHC.ppt

17:50 BaBar Online-Offline - Some Thoughts

A short overview of the BaBar Online/Offline interplay, lessons learned and some candidate topics for SuperB R&D

Speaker: Dr Steffen Luitz (SLAC)

Slides luitz-online-offline.pdf

18:05 CMS Online Software, Infrastructure Overview

The CMS online software encompasses the whole range of elements involved in the CMS data acquisition function. A central online software framework (XDAQ Cross-Platform DAQ Framework) has matured over a height years development and testing period and has shown to be able to cope well with the CMSrequirements. The framework relies on industry standard networks and processing equipment. All subsystems of the DAQ have adopted the central online software framework with the philosophy of using a common and standardized technology in order to reduce the effort associated with the maintenance and evolution of the detector read-out system over a long lifetime of the experiment. Adhering to a single software infrastructure in all subsystems of the experiment imposes a number of different requirements. High efficiency and configuration flexibility are among the most important ones. This presentation gives an overview of the software infrastructure, the architecture and design as well as brief outlook on the approaches and methods used to deal with the most technically relevant requirements.

Speaker: Luciano Orsini (CERN)

Slides SuperB-RD-WorkShop2010_V4.ppt

18:35 Discussion

20:30 → 22:30 Workshop Dinner

Thursday, 11 March

08:30 → 11:00 Plenary Session: Databases

08:30 Introduction, goals and an initial question

Speaker: I Gaponenko (SLAC)

Slides DatabasesIntro.pdf DatabasesIntro.ppt

08:45 Databases in BABAR

BABAR has been a unique experiment in a history of the High Energy Physics and not only due to the Physics itself, but also in a way this physics was extracted. It pioneered in involving hundreds of physicists, engineers and students into a practical C++ programming using OOAD methodologies. The experiment introduced a geographically distributed data processing and events simulation production system. Finally, it was the first (and so far - the only) experiment in a history of HEP which relied upon a commercial Object-Oriented Database to store detector events. At some point, almost every single bit of data, including detector events, were stored in databases. From the very beginning of the Experiment, databases were an integral part of the BABAR Computing Model. Over a course of the 9 years of data taking, our understanding on what are the right applications for databases and how they should be used had dramatically evolved. The experiment is now over, but its legacy still does matter. This talk is about our practical experience with various database technologies, a genesis of our understanding of how to design, implement and maintain a very complex system depending in its foundation on databases. It's about our successes and failures. Certain software artifacts of the experiment may also be found directly usable for the future Super-B.

Speaker: I. Gaponenko

Slides BabarDatabases.pdf BabarDatabases.ppt

09:15 Databases in LHC experiments: usage and lessons learned

Database applications and computing models of LHC experiments. Database integration with frameworks and into an overall data acquisition /production/processing and analysis chain of the experiment(s). Best practices in database development cycle: schema, contents, code deployment, etc. Choice of database technologies: RDBMS, hybrid, etc. Examples of database interfaces and tools, such as custom database interfaces and services on top of CORAL/COOL, etc. Scalability of database access in the distributed computing environment of LHC experiments. Distributed database applications: requirements, data complexity, update frequency, data volumes, usage, etc. Database areas in which substantial progresses was made and select remaining problems.

Speaker: A Vaniachine (Argonne National Laboratory)

Slides Databases_in_LHC_Experiments.pdf

10:00 Discussion

11:00 → 11:30 Coffee Break

11:30 → 13:00 Break-out groups working session

13:00 → 14:30 Lunch

14:30 → 17:00 Plenary Session - Distributed Computing

This session explores current status and future possibilities of distributed computing infrastructure for HEP experiments

14:30 Introduction and goals

Speakers: Armando Fella (CNAF), Eleonora Luppi (Ferrara University & INFN)

Slides intro_distributed_session.odp intro_distributed_session.pdf

14:40 Distributed Computing for HEP - present and future

In this presentation, the main aspects related to distributed computing that a HEP experiment has to address are discussed. This is done analyzing what the current experiments, mainly at the CERN LHC, are using, either provided by Gird infrastructures or developed by themselves. After a brief introduction on the overall distributed computing architecture, the specific aspects that are treated are: security and access control, data and storage management, workload management (with particular attention to user analysis) and infrastructure management.

Speaker: Claudio Grandi (INFN-Bologna)

Slides Workshop-SuperB-100311.pdf Workshop-SuperB-100311.ppt

15:10 Middleware development in the EGI era

After the end of the EGEE series of projects, the way improvements to the middleware are developed and deployed will change significantly, with a stronger focus on stability of the infrastructure and on quality assurance. The middleware will evolve according to the requirements coming from users in terms of reliability, usability, functionality, interoperability, security, management, monitoring and accounting, also exploiting emerging developments like virtualization. The presentation will briefly introduce the organizational framework for the middleware management in the EGI era, followed by an overview of the major expected developments.

Speaker: F. Giacomini (INFN - CNAF)

Slides mw_development_egi.odp mw_development_egi.pdf

15:30 Multithread and MPI usage in GRID

The talk will be focused on the support for parallel programs, mpi and multithread, in EGEE, that is actually the Grid infrastructure of which InfnGrid is member of. I’ll review the status of the support of MPI and the usage of the upcoming new syntax through the analysis of a Case Study.

Speaker: R. Alfieri (Parma University & INFN)

Slides mpi-grid.pdf

15:50 Grid/Cloud computing evolution

The presentation will first describe what the main current evolution trends for distributed computing seem to be, moving then on to explore how computing resources could be uniformly accessed via either grid or cloud interfaces using virtualization technologies. The integration of grids and clouds can on the one hand expand and optimize the use of available computing resources, while on the other hand allowing new scenarios for the solution of distributed computing tasks.

Speaker: Salomoni, Davide (INFN - CNAF)

Slides salomoni_superb_fe_100311.pdf

16:10 Discussion

17:00 → 17:30 Coffee Break

17:30 → 19:00 Plenary Session: User tools and interfaces

17:30 Introduction and goals

Speaker: Fabrizio Bianchi (TO)

Slides BianchiUserTools.pdf

17:40 Critical aspects of user interaction with the analysis environment

Using experience gained from the analysis environment in LHCb, BaBar and ATLAS, I will comment on the critical aspects that can make an analysis environment effective. I will do this by creating a set of requirements and then look at what lessons can be learned from past and current analysis environments with respects to these requirements. Some particular issues to consider are: - The unwanted separation between users and developers of analysis tools - Integration of user datasets into databases - Duplication of data (Ntuples that are straight copies of DSTs) - Software training - Analysis in a distributed environment

Speaker: Ulrik Egede (Imperial College (London))

Slides SuperBComputingMar2010.pdf

18:10 Visualizations and interfaces for end users

* A discussion of 4vector viewers as a visualization tool residing between the event display and final histograms. This has been useful for educating new students/analysts as well as an outreach tool. * Viewpoints: a NASA developed mult-variate display package used to develop more of an intuition for multi-variate datasets. This has been used with great success with undergraduate students but should not be limited to that level of experience. * A few comments about the revival of the 30-yo LASS data and what we might learn from that experience with respect to data format design.

Speaker: Mat Bellis

Slides Bellis_SuperB_Computing_Workshop.pdf

Video

• BtoDpi.mpg.mpg
• Ddecay_with_massplot.mpg.mpg
• viewpoints_usage.mpg

18:40 Discussion

Friday, 12 March

08:30 → 10:30 Plenary Session: Performance and efficiency of large data storage

08:30 Introduction and Goals

Speakers: Dr Fabrizio Furano (CERN), Vincenzo Maria Vagnoni (BO)

Slides StorageIntroduction.pdf

08:40 Overview and outlook of existing data access solutions

Speakers: Dr Fabrizio Furano (CERN), Vincenzo Maria Vagnoni (BO)

Slides OverviewStorage.pdf OverviewStorage.pptx

09:10 Research, Development and Scientific Application of Gfarm File System

Gfarm File System is a wide-area distributed file system that federates local disks of compute nodes in a Grid or computer clusters. It is a high performance distributed parallel file system designed for I/O intensive scientific data analysis conducted in collaboration with multiple distant organizations. Gfarm is under vigorous development for better performance and usability. A parallel distributed workflow system is developed to make the most use of Gfarm features such as file affinity scheduling. A case study of large-scale data analysis in astronomy using Gfarm file system shows a good scalable performance. In this talk we review recent researches and development of Gfarm.

Speaker: Masahiro Tanaka

Slides SuperB-Gfarm.pdf

09:35 Overview of the new technologies and evolution of storage systems for handling large volume of data

A report on some new emerging technologies suitable for managing, with high edfficiency, huge amount of data will be given. In particular details about Lustre, Hadoop and Ceph, as examples of different approaches to tackle the problem of providing input/output data to scientific applications, will be presented.

Speaker: Giacinto Donvito (INFN)

Slides donvito_storage_technologies_superB-WS-03-2010.pdf

10:05 Simulating storage system performance: a useful approach for SuperB ?

Capacity planning is a very useful tool to estimate future resource demand to carry on a given activity. Using well known modelling techniques it is possible to study the performance of a system before actually building it, and evaluate different design alternatives. However, capacity planning must be done properly in order to be effective. This presentation describes the benefits, requirements and pitfalls of performance modelling with a particular emphasis on storage systems.

Speaker: Dr Moreno Marzolla (Bologna University)

Slides Marzolla_superB_sim_2010.odp Marzolla_superB_sim_2010.pdf

10:30 → 11:00 Coffee Break

11:00 → 12:00 Plenary Session: Performance and efficiency of large data storage (II)

11:00 Discussion

Slides Discussion.pptx

12:00 → 13:00 Break out groups working session

13:00 → 14:30 Lunch

14:00 → 16:30 Plenary Session

Reports from parallel sessions on proposed goals, activites and resources needed for the SuperB R&D program; closeout

14:00 R&D proposal from the sessions "Impact of new CPU architectures, software architectures and frameworks"

Speakers: Peter Elmer (Princeton Univ.), Vincenzo Innocente (CERN)

Slides superb_elmer.pdf

14:30 R&D proposal from the session "Code development: languages, tools, standards and QA"

Speakers: Andrea Di Simone (RM2), Prof. Roberto Stroili (Universita' di Padova and INFN), Dr Steffen Luitz (SLAC)

Slides Coding-QA-Deployment.pdf

14:45 R&D proposal from the session "Persistence, data handling models and databases"

Speakers: Dr David Brown (Lawrence Berkeley National Lab), Sasha Vanyashin (CERN)

Database R&D Database_R&D.pdf

Slides Data_summary.pdf

15:05 R&D proposal from the session "Distributed Computing"

Speakers: Armando Fella (CNAF), Eleonora Luppi (Ferrara University & INFN)

Slides RD_proposal_distributed_comp.odp RD_proposal_distributed_comp.pdf

15:25 R&D proposal from the session "User tools and interfaces"

Speaker: Fabrizio Bianchi (TO)

Slides R&DProjects.pdf

15:45 R&D proposal from the session "Performance and efficiency of large data storage"

Speakers: Dr Fabrizio Furano (CERN), Vincenzo Maria Vagnoni (BO)

Slides Discussion.pdf Discussion.pptx

16:05 Discussion