You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: doc/Sphinx/Use/GPU_version.rst
+2-4
Original file line number
Diff line number
Diff line change
@@ -16,8 +16,6 @@ This page contains the links of this documentation to compile and run SMILEI on
16
16
17
17
----
18
18
19
-
Known issues
20
-
^^^^^^^^^^^^
19
+
Important note:
21
20
22
-
2D and 3D runs may crash with A2000 & A6000 GPUs (used in laptops and worstations respectively,
23
-
they are not 'production GPUs' which are designed for 64 bits floating point operations )
21
+
The biggest challenge to execute SMILEI on an accelerator is the correct installation of the openmpi library. It needs to be compiled with nvc++ after configuring (ie. ./configure --options) with the appropriate options specific to your system
Copy file name to clipboardExpand all lines: doc/Sphinx/Use/install_linux_GPU.rst
+20
Original file line number
Diff line number
Diff line change
@@ -6,6 +6,9 @@ First, make sure you have a recent version of CMAKE, and the other libraries
6
6
to compile Smilei on CPU as usual. In particular, for this example, you
7
7
need GCC <= 12.
8
8
9
+
The installation protocol showed below uses the openmpi included in nvhpc. This approach often results in segfault at runtime (note that nvidia will remove openmpi from nvhpc in the future).
10
+
The "proper" way, which is much harder, consists in installing openmpi compiled with nvhpc (
11
+
9
12
Make a directory to store all the nvidia tools. We call it $NVDIR:
10
13
11
14
.. code:: bash
@@ -72,3 +75,20 @@ To run:
72
75
73
76
source nvidia_env.sh
74
77
smilei namelist.py
78
+
79
+
80
+
As an example of a "simple" openmpi installation
81
+
Openmpi dependencies such as zlib, hwloc and libevent should first be compiled with nvc++
0 commit comments