en-es
en-fr
en-sl
en
0.25
0.5
0.75
1.25
1.5
1.75
2
Life Science: Using HPC for Insight into Biomolecular Function
Published on Sep 19, 20161050 Views
Related categories
Chapter list
Life Science: Using HPC for Insight into Biomolecular Function00:00
Intro02:03
Why multiple life sciences lectures? - 105:08
Why multiple life sciences lectures? - 205:49
Why multiple life sciences lectures? - 306:24
Thomas Cheatham, III 06:41
Erik Lindahl09:23
What do we want to do?12:16
Accurate modeling of molecules requires12:58
What is a molecular mechanical "force field"?15:28
Molecular simulation / molecular dynamics16:46
Michael Levitt16:51
Molecular simulation17:36
Challenge 118:18
An example of code evolution - 119:56
An example of code evolution - 219:56
An example of code evolution - 320:15
An example of code evolution - 421:47
An example of code evolution - 522:30
…evolution of AMBER 4.1 codes - 124:21
…evolution of AMBER 4.1 codes - 224:48
…evolution of AMBER 4.1 codes - 326:17
…evolution of AMBER 4.1 codes - 426:50
AMBER 16 (released ~May 2016)26:52
Challenge 227:17
How to fully sample conformational ensemble? - 127:48
How to fully sample conformational ensemble? - 227:59
Convergence, force field and salt dependence in simulations of nucleic acids28:37
How to test for convergence between two simulations?28:56
Test for convergence within and between simulations... - 130:50
Test for convergence within and between simulations... - 231:26
If we cannot scale to larger machine (more cores), couple independent MD simulations: i.e., use ensembles (replica exchange, || tempering, Markov State modeling, …)31:41
All MPI communications32:36
CommWorld33:06
CommSander - 133:14
CommSander - 233:21
Create a communicator for each group of -ng NumGroup processors33:27
Define a communicator33:29
r(GACC) tetranucleotide34:59
Other issues35:48
multi-D REMD35:49
H-REMD & M-REMD - 135:57
CPPTRAJ in AmberTools36:12
H-REMD & M-REMD - 237:58
Problems with our tightly coupled approach - 138:13
Problems with our tightly coupled approach - 238:34
Setup, analysis, data management, … - 139:57
Setup, analysis, data management, … - 242:13
Automate analysis & tools for deeper “interactive” analysis - 142:53
Automate analysis & tools for deeper “interactive” analysis - 243:55
Peta- or exa- scale science: the problem will only get worse!43:56
Data challenges - 145:55
Data challenges - 246:55
Life Science: How Do We Solve Problems Faster in Parallel and on Accelerators?49:48
Costly, because these terms involve all pairs56:22
The challenge57:07
Historical approaches to make our codes faster57:59
Example: Remove FLOPS by taking longer steps59:19
Example: Remove FLOPS by using smaller simulation boxes59:28
How do we find parallelism?01:00:00
What does a modern CPU look like?01:01:42
Execute 4 iterations of the innermost loop at once01:02:28
Explicit Data Parallelism01:04:12
Hardware01:04:41
It is much easier to port and scale a simple reference program01:05:42
A failed GPU attempt?01:06:05
Option 1: Stay on the GPU01:07:13
AMBER01:08:15
Hardware prediction01:09:37
Domain decomposition dynamic load balancing01:10:36
Heterogeneous CPU-GPU acceleration in GROMACS01:11:09
You cannot use neighborlists…01:12:12
Tiling circles is difficult01:13:01
From neighborlists to cluster proximity lists01:13:59
Unified GPU/CPU architecture - completely portable01:14:46
Surprisingly little CUDA code01:15:14
Kernel timing01:15:34
CUDA overhead & scaling issues01:15:42
A lot of low-level tuning01:16:13
Integrating the GPU cont.01:17:13
The Villin headpiece01:17:15
Desktop example: Core i7 4790K & GTX Titan01:17:43
Strong scaling01:18:13
You have already lost the CPU game - 101:18:17
You have already lost the CPU game - 201:18:27
You have already lost the CPU game - 301:18:33
No longer true: 14nm transition01:18:34
Memento mori01:19:08
How will YOU use a billion cores?01:19:41
Simulating the whole cell01:21:01
End result is exactly the same01:21:04
From ~100k cores to Exascale: Ensembles01:21:10
Markov State Models - 101:21:42
Markov State Models - 201:21:46
Markov State Models - 301:21:54
Markov State Models - 401:21:56
Markov State Models - 501:22:01
Markov State Models - 601:22:05
Ensembles in action01:22:39
30 hours later01:22:40
Thank you01:22:49