jianmu
/
lammps
forked from lijiext/lammps
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

git-svn-id: svn://svn.icms.temple.edu/lammps-ro/trunk@3060 f3b2605a-c512-4ea7-a41b-209d697bcdaa

This commit is contained in:
sjplimp 2009-08-13 17:12:38 +00:00
parent ea26344b52
commit bff67c37d7
2 changed files with 146 additions and 50 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

98
doc/Section_start.html
View File

@ -748,35 +748,83 @@ communication, roughly 75% in the example above.
<H4><A NAME = "2_8"></A>2.8 Running on GPUs
</H4>
<P> 1. I assume you have nvidia card ( cat /proc/driver/nvidia/cards/0 )
2. Go to http://www.nvidia.com/object/cuda_get.html
3. Install latest driver and toolkit (SDK is not necessary)
4. Run make in lib/gpu
5. ./nvc_get_devices will list supported devices and properties
<P>A few LAMMPS <A HREF = "pair_style.html">pair styles</A> can be run on graphical
processing units (GPUs). We plan to add more over time. Currently,
they only support NVIDIA GPU cards. To use them you need to install
certain NVIDIA CUDA software on your system:
</P>
<P>The pair styles supporting GPU acceleration have optional arguments to
determine how GPUs are selected. Each GPU is selected based on its ID
number reported by the graphics driver. A list of supported GPUs can be
obtained using the 'nvc_get_devices' executable that is built with the
library. To change how GPUs are selected, a GPU selection mode and a GPU
selection parameter are supplied. If the mode is "one_node", the GPU ID
is equal to the process rank plus the GPU parameter. For example,
<UL><LI>Check if you have an NVIDIA card: cat /proc/driver/nvidia/cards/0
<LI>Go to http://www.nvidia.com/object/cuda_get.html
<LI>Install a driver and toolkit appopriate for your system (SDK is not necessary)
<LI>Run make in lammps/lib/gpu, editing a Makefile if necessary
<LI>Run lammps/lib/gpu/nvc_get_devices to list supported devices and properties
</UL>
<H4>GPU hardware
</H4>
<P>When using GPUs, you are restricted to one physical GPU per LAMMPS
process. This can be multiple GPUs on a single node or across
multiple nodes. For each GPU pair style, the first two arguments (GPU
mode followed by GPU parameter) control how GPUs are selected. If you
are running on a single node, the mode is "one/node" and the parameter
is the ID of the first GPU to select:
</P>
<P>pair_style lj/cut/gpu 2.5 one_node 0
<PRE>pair_style lj/cut/gpu one/node 0 2.5
</PRE>
<P>The ID is the GPU ID reported by the driver for CUDA enabled graphics
cards. For multiple GPU cards on a node, an MPI process should be run
for each graphics card. In this case, each process will grab the GPU
with ID equal to the process rank plus the GPU parameter.
</P>
<P>will select GPUs 0, 1, and 2 for a 3 process run. If the mode is
"one_gpu", every process will select the same GPU id identified by the
GPU parameter. This option can be used to select a single GPU on
multiple nodes and should be run with exactly 1 process per node. If the
mode is "multi_gpu" the GPU parameter supplied should be the number of
GPUs per node. The GPU selected by a particular process will then be
equal to the rank%gpus_per_node. This final option can be used to run on
multiple nodes that contain multiple GPUs per node. It assumes that
processes will be filled in order by slots such that the ranks on any
given node are always sequential.
<P>For multiple nodes with one GPU per node, the mode is "one/gpu" and
the parameter is the ID of the GPU used on every node:
</P>
<P>double/single precision
atom limits
<PRE>pair_style lj/cut/gpu one/gpu 1 2.5
</PRE>
<P>In this case, MPI should be run with exactly one process per node.
</P>
<P>For multiple nodes with multiple GPUs, the mode is "multi/gpu" and the
parameter is the number of GPUs per node:
</P>
<PRE>pair_style lj/cut/gpu multi/gpu 3 2.5
</PRE>
<P>In this case, LAMMPS will attempt to grab 3 GPUs per node and this
requires that the number of processes per node be 3. The first GPU
selected must have ID zero for this mode (in the example, GPUs 0, 1,
and 2 will be selected on every node). An additional constraint is
that the MPI processes must be filled by slot on each node such that
the process ranks on each node are always sequential. This is a option
for the MPI launcher (mpirun/mpiexec) and will be the default on many
clusters.
</P>
<H4>GPU single vs double precision
</H4>
<P>See the lammps/lib/gpu/README for instructions on how to build the
LAMMPS gpu library for single vs double precision. The latter
requires that your GPU card supports double precision.
</P>
<H4>GPU Memory
</H4>
<P>Upon initialization of the pair style, LAMMPS will reserve memory for
64K atoms per GPU or 70% of each card's GPU memory, whichever value is
limiting. If the GPU library is compiled for double precision, the
maximum number of atoms per GPU is 32K. When running a periodic
system and/or in parallel, this maximum atom count includes ghost
atoms.
</P>
<P>The value of 70% can be changed by editing the PERCENT_GPU_MEMORY
definition in the appopriate lammps/lib/gpu source file. The value of
64K cannot be increased and is the maximum number of atoms allowed per
GPU. By default, enough memory to store at least the maximum number
of neighbors per atom is reserved on the GPU, which is set by the
<A HREF = "neigh_modify.html">neigh_modify one</A> command. The default value of
2000 will be very high for many cases. If memory on the graphics card
is limiting, the number of atoms allowed can be increased by
decreasing the maximum number of neighbors. For example placing,
</P>
<PRE>neigh_modify one 100
</PRE>
<P>in the input script will decrease the maximum number of neighbors per
atom to 100, allowing more atoms to be run on the GPU.
</P>
<HR>

98
doc/Section_start.txt
View File

@ -741,35 +741,83 @@ communication, roughly 75% in the example above.
2.8 Running on GPUs :h4,link(2_8)
1. I assume you have nvidia card ( cat /proc/driver/nvidia/cards/0 )
2. Go to http://www.nvidia.com/object/cuda_get.html
3. Install latest driver and toolkit (SDK is not necessary)
4. Run make in lib/gpu
5. ./nvc_get_devices will list supported devices and properties
A few LAMMPS "pair styles"_pair_style.html can be run on graphical
processing units (GPUs). We plan to add more over time. Currently,
they only support NVIDIA GPU cards. To use them you need to install
certain NVIDIA CUDA software on your system:
The pair styles supporting GPU acceleration have optional arguments to
determine how GPUs are selected. Each GPU is selected based on its ID
number reported by the graphics driver. A list of supported GPUs can be
obtained using the 'nvc_get_devices' executable that is built with the
library. To change how GPUs are selected, a GPU selection mode and a GPU
selection parameter are supplied. If the mode is "one_node", the GPU ID
is equal to the process rank plus the GPU parameter. For example,
Check if you have an NVIDIA card: cat /proc/driver/nvidia/cards/0
Go to http://www.nvidia.com/object/cuda_get.html
Install a driver and toolkit appopriate for your system (SDK is not necessary)
Run make in lammps/lib/gpu, editing a Makefile if necessary
Run lammps/lib/gpu/nvc_get_devices to list supported devices and properties :ul
pair_style lj/cut/gpu 2.5 one_node 0
GPU hardware :h4
will select GPUs 0, 1, and 2 for a 3 process run. If the mode is
"one_gpu", every process will select the same GPU id identified by the
GPU parameter. This option can be used to select a single GPU on
multiple nodes and should be run with exactly 1 process per node. If the
mode is "multi_gpu" the GPU parameter supplied should be the number of
GPUs per node. The GPU selected by a particular process will then be
equal to the rank%gpus_per_node. This final option can be used to run on
multiple nodes that contain multiple GPUs per node. It assumes that
processes will be filled in order by slots such that the ranks on any
given node are always sequential.
When using GPUs, you are restricted to one physical GPU per LAMMPS
process. This can be multiple GPUs on a single node or across
multiple nodes. For each GPU pair style, the first two arguments (GPU
mode followed by GPU parameter) control how GPUs are selected. If you
are running on a single node, the mode is "one/node" and the parameter
is the ID of the first GPU to select:
double/single precision
atom limits
pair_style lj/cut/gpu one/node 0 2.5 :pre
The ID is the GPU ID reported by the driver for CUDA enabled graphics
cards. For multiple GPU cards on a node, an MPI process should be run
for each graphics card. In this case, each process will grab the GPU
with ID equal to the process rank plus the GPU parameter.
For multiple nodes with one GPU per node, the mode is "one/gpu" and
the parameter is the ID of the GPU used on every node:
pair_style lj/cut/gpu one/gpu 1 2.5 :pre
In this case, MPI should be run with exactly one process per node.
For multiple nodes with multiple GPUs, the mode is "multi/gpu" and the
parameter is the number of GPUs per node:
pair_style lj/cut/gpu multi/gpu 3 2.5 :pre
In this case, LAMMPS will attempt to grab 3 GPUs per node and this
requires that the number of processes per node be 3. The first GPU
selected must have ID zero for this mode (in the example, GPUs 0, 1,
and 2 will be selected on every node). An additional constraint is
that the MPI processes must be filled by slot on each node such that
the process ranks on each node are always sequential. This is a option
for the MPI launcher (mpirun/mpiexec) and will be the default on many
clusters.
GPU single vs double precision :h4
See the lammps/lib/gpu/README for instructions on how to build the
LAMMPS gpu library for single vs double precision. The latter
requires that your GPU card supports double precision.
GPU Memory :h4
Upon initialization of the pair style, LAMMPS will reserve memory for
64K atoms per GPU or 70% of each card's GPU memory, whichever value is
limiting. If the GPU library is compiled for double precision, the
maximum number of atoms per GPU is 32K. When running a periodic
system and/or in parallel, this maximum atom count includes ghost
atoms.
The value of 70% can be changed by editing the PERCENT_GPU_MEMORY
definition in the appopriate lammps/lib/gpu source file. The value of
64K cannot be increased and is the maximum number of atoms allowed per
GPU. By default, enough memory to store at least the maximum number
of neighbors per atom is reserved on the GPU, which is set by the
"neigh_modify one"_neigh_modify.html command. The default value of
2000 will be very high for many cases. If memory on the graphics card
is limiting, the number of atoms allowed can be increased by
decreasing the maximum number of neighbors. For example placing,
neigh_modify one 100 :pre
in the input script will decrease the maximum number of neighbors per
atom to 100, allowing more atoms to be run on the GPU.
:line