... jgopt.org ...

| Home | Downloads | Legal |

Proudly supported by:


Current Status (6-September-2013):
X-Branch (Cuda) development history, structure & imlementation is being reviewed, for generalisation into a heterogeneous design (Superdreadnaught)

Current Projects:

X-Branch (Active, design review)
AppMan
XsigLib
ModBoinc
Superdreadnaught

seti@home Cuda multibeam (X-branch)
Codename: X-Branch
Project Status: Detailed review of initial public release, version x41zc, seti@home project
Brief Description: Developed from Berkeley 6.09 application sources, refined to extend support for newer Cuda enabled hardware, for detailed study & experimentation. All gpGPU supercompute related issues & ongoing platform evolution (hardware, drivers, techniques). Establish known issues for further designs. Use results to study GPU and cross platform validation and reliability issues, toward design of new domain specific end user tools for assessing, diagnosing, tuning & comparing equipment & software, as well as end-user and developer practices and recommendations.
Downloads: Current version (x41zc) Distributed as seti@home stock appication, and via the third party Lunatics Unified Installer
Participation: Sources available via Berkeley's svn repository at https://setisvn.ssl.berkeley.edu/svn/branches/sah_v7_opt/Xbranch , email contact at jgopt dot org

Boinc 3rd party Application Manager
Codename: AppMan
Project Status: Detailed design stage
Brief Description: The need for simplified gui based boinc science application management & configuration under the Boinc Platform, along with recent compiler/library licensing issues exposed limitations in the evolved form of managing Boinc applications under anonymous platform, namely maintenance overhead associated with unified package release in the event of rapid development changes, or other changes in circumstances. A secured, self updating & maintaining system is proposed to address these limitations, based loosely on schemes used by popular online gaming engines, on a cross project, cross development team footing.
Participation: informal whitepaper & proof of concept / prototypes under development, intended for user and developer comment. contact at jgopt dot org

Open Source libraries for signal processing, and moves to free toolchains for seti@home CPU based applications
Codename: XsigLib
Project Status: Research & prior experiment consolidation
Brief Description: Historic work by Lunatics members and Berkeley contributors utilised a wide variety of hand coded design/compile time and runtime optimisation technologies. These are associated with the use of profiling/runtime dispatch techniques, hard coded microarchitectural optimisations at compile time, adaptive dynamic runtime techniques such as those used by the fftw Fourier transform library, and associated vendor libraries such as Intel's Integrated performance primitives Spiral project based work, and hand assembly coding techniques. This project seeks to firstly rationalise the disparate territories. Libraries will be adding current research based off-line (install-time) optimisation techniques, as well as extending search techniques into specific problem domain knowledge related optimisation strategies (e.g. prior 'Cascade Freaky powerspectra' experiments) as extension & substitutes for the multiple chirprate and fold level STFT used currently in all seti@home multibeam narrowband search applications. Techniques for entropy guided search area exclusion will be examined in chirplet & foldlet space. It is expected that these libraries can be generalised for use in many fields using multiresolution signal analysis in Fourier Power over time, Frequency-Time, chirplet or foldlet space as specialised extensions of wavelet space. Prior work from Berkeley, Lunatics members, Alex Kan and other contributors in combination with recent research into off-line optimisation extensions to adaptive runtime dynamic optimisation will be examined. Ultimate goal is a cross platform library from scratch under open source licence, such that algorithms can be specified & guide design-time, offline(install-time), and adaptive runtime optimisation with the presumption of changing environmental conditions, particulary 'self optimising' across traditional single threaded through heterogeneous local and remote clustering.
Participation: contact at jgopt dot org

Boinc client & server development
Codename: ModBoinc
Project Status: Under review for applicability under the current Boinc infrastructure
Brief Description: Experiments prior to extensive Boinc server side changes, revealed close ties between client task estimate and work fetch, with charateristic oscillations and convergence failures associated with task cache maintenance failures of various forms. The server side changes to work elapsed runtime estimates scaling embedded via task parameters largely alleviated medium to long term problems associated with innacurate work complexity estimation, though did not address short term response/convergence under data dependant variation, or under client side machine usage pattern changes. Under planned future heterogeneous designs both the existing medium to long term estimate convergence, and so work fetch character, should be responsive to short term runtime usage/environment/resource change, as well provide sufficient damping to prevent overshoot & ringing in the estimates characteristic of an untuned model. Prior experiments implemented a simple tuned PID controller on a per application basis as an additoinal local per application duration correction factor. Future work for heterogeneous environments seeks to migrate & refine this work to use the existing flops mechanisms, while providing a careful new system 'on-ramp' and run-time adaptability to resource change.
Participation: contact at jgopt dot org

Hetergeneous & LAN clustering
Codename: Superdreadnaught
Project Status: Long term technological convergence, research & design
Brief Description: Projected data processing needs for existing and future data acquisistion place task elapsed runtimes using traditional single threaded applications from a few times, to orders of magnitude longer than existing data searches. As compute performance advances already place existing searches under high levels of memory subsystem contention, it is expected that future search applications will benefit greatly from exploiting spatial and temporal data locality across more parallelised implementations. As multicore processors, GPU processing and household networks are available to increasing numbers of Boinc platform users, at some threshold mutithreaded heterogeneous designs will likely have increasing efficiency benefits over traditional single-threaded approaches. Convergence of evolving cross platform precision, reliability, design-time, offline & runtime adaptability are all aspects that will require extensive exploration, along with new models & protocol designs to cope with task domain specific & algorithmically aware scheduling.
Participation: contact at jgopt dot org

All contents herein are Copyright © Jason Groothuis, 2012, All Rights Reserved.