LibGeoDecomp 0.4.0 and LibFlatArray 0.2.0 Released 2014-10-27T00:00:00+01:00
Today we are proudly shipping two new releases of our
libraries LibGeoDecomp and LibFlatarray. These releases
are our first to deliver peta-scale performance on the
world's fastest supercomputers.
Together they represent the culmination of ten months
of work and more than 800 commits of an alliance of
from the Friedrich-Alexander-Universität
and Louisiana State
Google Summer of Code 2014 2014-03-18T00:00:00+01:00
The Ste||ar group has been accepted as a mentoring
organization in the Google Summer of Code 2014. If you'd
like to get involved in bleeding edge HPC research, and
even get a nice paycheck for it, take a look
LibFlatArray 0.1.0 Released 2014-01-23T00:00:00+01:00
Our daughter project LibFlatArray
, a super-fast 2D/3D container class, has shipped its inaugural release 0.1.0.
LibGeoDecomp hits 9.1 PFLOPS on Titan 2014-01-16T00:00:00+01:00
A short-ranged n-body simulation is not only the first
application built on LibGeoDecomp to break the PFLOPS
barrier, but continued to rush on towards 16384 nodes.
LibGeoDecomp 0.3.1 Released 2014-01-11T00:00:00+01:00
0.3.1 is a service release which bundles the missing
code of LibFlatArray and fixes some installer issues.
LibGeoDecomp 0.3.0 Released 2013-11-19T00:00:00+01:00
Today we are releasing LibGeoDecomp, version 0.3.0
. This release is the culmination of
eight months of work, consisting of over 700 commits and 9000 lines
of new code. We have added a plethora of new features, the most
exciting ones being support for Blue Gene/Q, the Xeon Phi (via the
new HPX backend), a revamped user code API, and an alternative struct
of arrays (SoA) memory layout.
6M Core-Hours Allocation on JUQUEEN 2013-10-09T00:00:00+02:00
The John von Neumann Institute for Computing (NIC) has
aproved our project proposal for the JUQUEEN supercomputer at the
Forschungszentrum Jülich. Together with the STEllAR group
we will work
on Auto-tuned Large-scale Simulations on BG/Q with
LibGeoDecomp and HPX
LibGeoDecomp at SC13 2013-10-09T00:00:00+02:00
This November we'll have a much more massive presence at the
SupercomputingConference (SC13, Denver, CO) than last year. And
we're really excited about this. LibGeoDecomp will be featured in
three talks in the technical program and one booth on the exhibits
floor. And then there is TeamKrautComputing...
How to Parallelize Fortran Kernels with LibGeoDecomp 2013-09-19T00:00:00+02:00
Recently the question popped up on Slashdot
whether you use LibGeoDecomp in Fortran. I took a look at it and, to
cut a long story short, the answer is: yes, it's surprisingly easy,
but you lose some features. Here is the long story:
LibGeoDecomp now licensed under Boost Software License 2013-08-17T00:00:00+02:00
To ease adoption of LibGeoDecomp in
commercial/confidential environments, we have decided to transition
from LGPL3 to the Boost Software License.
CFD Videos of LBM Toolkit Added to Gallery 2013-04-05T00:00:00+02:00
For his master's thesis Stephan Helou has written a LibGeoDecomp-based
toolkit for computational fluid dynamics (CFD) simulations based on
the Lattice Boltzmann method (LBM). I've added a collection of videos
to the gallery
LibGeoDecomp-related Project Ideas for GSOC 2013 2013-03-29T00:00:00+01:00
LibGeoDecomp is applying as a mentoring organization
Summer of Code 2013
. Below you'll find a list of
ideas for projects. Take these as rough guidelines.
You're most welcome to come to our IRC
channel or mailing
to discuss your own ideas.
A Glimpse Into the Future of GPGPUs 2013-03-19T00:00:00+01:00
Jen-Hsun Huang is currently delivering the keynote speech of the GPU Technology Conference 2013. He shared some news on Nvidia's roadmap, epecially on Volta, Logan, and Parker.
LibGeoDecomp 0.2.0 Released 2013-03-13T00:00:00+01:00
is now ready for general
consumption. This release focuses on scalability and robustness. Release slides are here
. Short changelog:
The GCC and Vectorization 2013-03-11T00:00:00+01:00
Stencil codes are often seen as a prime example for
real-world problems where vectorization can be applied
easily. After all, the same operations have to be
carried out for each grid cell and many prominent
kernels, e.g. LBM (Lattice
) or the RTM (Reverse
) don't even contain conditionals. So
it is even more surprising, that compilers still
struggle at generating vectorized code automatically. Or are times changing?
Infrastructure Updated 2013-02-02T00:00:00+01:00
Our project has made significant progress in the past
months. To allow us to keep up the pace I've updated
some of our infrastructure. The most promiment feature
are probably the automated performance tests
LibGeoDecomp at GTC 2013 2013-01-04T00:00:00+01:00
We're exited to anounce that we'll give a talk at
2013 in San Jose. In our presentation we'll show how to build a
small demo app which showcases a couple of LibGeoDecomp's new features:
utilization of CUDA capable GPUs, live steering, and in
LibGeoDecomp at SC12 2012-11-09T00:00:00+01:00
The scene is headed to SLC, and so are we. There will be two oppotunities to learn about LibGeoDecomp: we'll give a talk at the SC12 workshop WOLFHPC
and -- even more exitingly -- LSU will host a demo which we've put together with our colleagues from the STE||AR Group
KONWIHR Funding Approved 2012-09-28T00:00:00+02:00
We're receiving funding from the Bavarian KONWIHR
initiative to optimize the support for n-body simulations in LibGeoDecomp.
Scaling on 480k Cores on Tsubame 2.0 2012-09-23T00:00:00+02:00
Now that the conference Facing the
is over and the results have
been published, I can finally upload them here. The image summarizes...
Big O Notation vs. Real Life 2012-09-05T00:00:00+02:00
One point where LibGeoDecomp really shines is its powerful geometry subsystem. This part handles the domain decomposition and can adapt to every conceivable partitioning scheme. The downside of this flexibility is that we have to store the coordinates of the cells allocated on a node somewhere. This is the task of the Region
class, which performs a run-length compression to save memory. For an approximately cubic region of n³
cells it needs O(n²)
space, which should be negligible compared to the O(n³)
bytes for the grid itself. Well, should
Building on Tsubame 2.0 2012-07-24T00:00:00+02:00
I'm currently prepping the builds for Tsubame 2.0
. The machine is currently ranked #14 on the Top 500
, the list of the world's fastest computers. Gotta hurry though, as Prof. Aoki
kindly arranged for a reservation of 1000 GPUs on Sunday.
LibGeoDecomp 0.1.0 Released 2012-06-27T00:00:00+02:00
The future of scientific computing begins today. At least according to Back to the Future
. And I have to agree. Today we release LibGeoDecomp 0.1.0.
New scaling results on RRZE's LiMa 2012-06-14T00:00:00+02:00
We're finally moving to the big machines. Yesterday I gathered the first timing data from our jobs on RRZE's LiMa
cluster. We ran DendSim3
with LibGeoDecomp in a Gustafson and Barsis
setup. This is also called weak scaling, as we increase the grid volume linearly with the number of cores. So far we have achieved an efficiency in excess of 94% while running on 768 cores.
Rake's dead, baby, Rake's dead! +short history of LibGeoDecomp 2012-05-30T00:00:00+02:00
What may sound like a step backward is actually huge leap forward for us. I've thrown out the fancy shmancy Ruby/Rake
stuff and switched the build system to pure CMake. The previous build system was really comfy: it was capable of parallel, out-of-source builds, could handle code generation, run all unit-tests, did not require the user to manually list all source files... I could probably list another dozen features that I loved about it, but
it had a fatal flaw:
Website updated 2012-05-15T00:00:00+02:00
Everything looks so different now, and yet the same. It all started with me just
wanting to add a secondary nav bar, which appeared to be so simple...
Hooray, we got news 2012-05-12T00:00:00+02:00
...or to be precise: I've added a news section. It may be empty
for now, but it's a news section nevertheless.