-
Notifications
You must be signed in to change notification settings - Fork 100
QUDA on Summit
maddyscientist edited this page May 30, 2024
·
2 revisions
When found compiling QUDA on Summit with gcc-12, there is an issue compiling Eigen. The resolution is to use binutils/2.40
(as found by Colin Morningstar). The module environment that compiles QUDA is
module load DefApps-2023
module load binutils/2.40
module load cuda/12.2.0
module load cmake/3.23.2
module load gcc/12.1.0
When running on Summit, it may be beneficial to enable some dslash policies that utilize host memory for reading/writing halos directly. The policies in question can be beneficial on Summit, since it has poor GPU Direct RDMA (GDR) performance, however, the NVLink connection to the host memory can offset this deficit. Set the environment variable QUDA_ENABLE_ZERO_COPY=1
to enables these policies. These policies are disabled by default to reduce autotuning overhead for systems that have performant GDR.