Some projects I've worked on at D. E. Shaw Research
D. E. Shaw Research does
scientific research in computational biochemistry and molecular
dynamics. (See an up-to-date
list of D. E. Shaw Research papers and publications in various
leading science and engineering journals and conferences)
David Shaw's 2006 talk at Stanford University's Computer Systems
Colloquium has both
which provide background, context and elaboration for our lab's work.
We have designed and built Anton 2 and previously,
the fastest biomolecular simulation machines in the world. Some papers
describing science done on Anton:
How Fast-Folding Proteins Fold (Science, 28 Oct 2011) reports the
results of many long atomic-level molecular dynamics simulations on
Anton, each run for periods ranging between 100 µs and 1 ms of
simulated time, that reveal a a set of common principles underlying the
folding of 12 structurally diverse proteins. The proteins range in size
from 10 to 80 residues, and include members of all three majoral
structural classes (α-helical, β sheet and mixed
α/β). Each protein was simulated near the melting
temperature (so both folding and unfolding events were likely). For
each protein, between one and four simulations were run for long enough
to observe at least 10 folding and 10 unfolding events. In aggregate,
this represents ˜8 ms of simulation with more than 400 folding and
unfolding events. The proteins were all observed to fold close to their
experimentally determined native structures.
How Does a Drug Molecule Find Its Target Binding Site?(JACS, 5 May
2011) provides a continuous, atomic-level view of binding process of a
ligand (the cancer drug dasatinib or the kinase inhibitor PP1) to a
protein (Src kinase). In long, unguided simulations on Anton, the
ligand, initially placed at a random location within a box that also
contained the protein, correctly identified its target binding site.
Watch the movie: the ligand is the orange-brown molecule.
Atomic-Level Characterization of the Structural Dynamics of Proteins
(Science, 15 Oct 2010) describes results of various record-length
molecular dynamics simulations on Anton, including a millisecond-long
simulation of BPTI (bovine pancreatic trypsin inhibitor) and hundreds of
microseconds of WW domain and villin headpiece, including complete
folding simulations from a fully extended state]
Papers describing Anton 2:
Papers describing Anton:
Extending the Generality of Molecular Dynamics Simulations on a
Special-Purpose Machine (IPDPS 2013, Boston, Best Paper,
Applications) describes how we achieved high performance on several
advanced molecular dynamics features.
Millisecond-scale molecular dynamics simulations on Anton
(SuperComputing 2009, Best Paper, Gordon Bell Prize for Special
Achievement) describes Anton's algorithms and reports measured
performance running the longest molecular dynamics simulations that
anyone has run (as far as we know).
Anton, a special-purpose machine for molecular dynamics simulation
(Communications of the ACM, July 2008)
162-Nanosecond End-to_end Communication Latency on Anton
describes the combination of hardware mechanisms and software constructs
that put Anton's communication network ahead of what's been achieved
on most other clusters and supercomputers.
Anton, a special-purpose machine for molecular dynamics simulation,
High-Throughput Pairwise Point Interactions in Anton, a Specialized
Machine for Molecular Dynamics Simulation (HPCA 2008).
Incorporating Flexibility in Anton, a Specialized Machine for Molecular
Dynamics Simulation (HPCA 2008).
A side-project that emerged from Anton is a class of "counter-based"
random number generators (CBRNGs), which we presented at SC11
(awarded Best Paper). CBRNGs are fast, stateless functions that are
great for modern multi-core or distributed applications -- the source
code is available for download. It was interesting getting them to work
in C, C++0X as well as OpenCL (on an AMD HD6970) and CUDA (on an NVIDIA
GTX580). NVIDIA includes one of these generators in their CURAND library,
starting with version 6, they've been described in books like
Numerical Computation with GPUS, and papers like
Random number generators for massively
parallel simulations on GPU. There's also a
Haskell port of Random123, and
a Julia implementation.
There's a version of
Random123 in Fujitsu's Open Petascale library.
GitHub has a Boost-compatible
implementation of Random123.
Another of our projects is Desmond, a scalable
molecular dynamics (MD) program for commodity clusters (aka
We first described the performance of Desmond on a 2005-vintage
Infiniband+Opteron cluster in
Scalable Algorithms for Molecular Dynamics Simulations on Commodity
Clusters, (SuperComputing 2006, Best Paper). We also released a
tech report with updated Desmond performance numbers on our 2nd generation
(2008-vintage) Infiniband+Xeon cluster, and Desmond performance numbers on modern NVIDIA GPUs.
I gave a
talk on D. E. Shaw Research's experience with deploying our first
generation (vintage 2005)
Infiniband+Opteron cluster at the
2007 OpenFabrics Sonoma Workshop
The D. E. Shaw Research publications page.
More general articles about D. E. Shaw Research:
My home page.