DokuWiki

We are interested in benchmarking the serial, MPI, cuda and cuda.MPI versions of pmemd.

• Verified the MPI threads and GPU invocations
• Verified the output data
• pmemd.cuda.MPI errors
• Script used is listed at end of this page

• MPI speedup near -np 24 is 8x serial

• GPU serial speedup is 17.5x CPU serial performance and outperforms MPI by at least 2x
• GPU parallel unable to measure

• 5-6x performance speed ups using one GPU versus 16 CPU cores
• 9-10x perrformance speedups using four GPUs versus 16 CPU cores

First we get some CPU based data.

# serial run of pmemd
nohup $AMBERHOME/bin/pmemd -O -i mdin -o mdout -p prmtop \
-c inpcrd -r restrt -x mdcrd </dev/null &

# parallel run, note that you will need create the machinefile
# if -np=4 it would would contain 4 lines with the string 'localhost'...does not work, use hostname
mpirun --machinefile=nodefile -np 4 $AMBERHOME/bin/pmemd.MPI \
-O -i mdin -o mdout -p prmtop \
-c inpcrd -r restrt -x mdcrd </dev/null &

The following script should be in your path … located in ~/bin

You need to allocate one or more GPUs for your cuda runs.

node2$ gpu-info
====================================================
Device  Model           Temperature     Utilization
====================================================
0       Tesla K20       27 C             0 %
1       Tesla K20       28 C             0 %
2       Tesla K20       27 C             0 %
3       Tesla K20       30 C             0 %
====================================================

Next we need to expose these GPUs to pmemd …

# expose one
export CUDA_VISIBLE_DEVICES="0"

# serial run of pmemd.cuda
nohup $AMBERHOME/bin/pmemd.cuda -O -i mdin -o mdout -p prmtop \
-c inpcrd -r restrt -x mdcrd </dev/null &

# parallel run, note that you will need create the machinefile
# if -np=4 it would could contain 4 lines with the string 'localhost'
mpirun --machinefile=nodefile -np 4 $AMBERHOME/bin/pmemd.cuda.MPI \
-O -i mdin -o mdout -p prmtop \
-c inpcrd -r restrt -x mdcrd </dev/null &

You may want to try to run your pmemd problem across multiple GPUs if problem set is large enough.

# expose multiple (for serial or parallel runs)
export CUDA_VISIBLE_DEVICES="0,2"

[TestDriveUser0@K20-WS]$ cat run
#!/bin/bash
rm -rf err out logfile mdout restrt mdinfo

echo CPU serial
pmemd -O -i inp/mini.in -p 1g6r.cd.parm \
 -c 1g6r.cd.randions.crd.1 -ref 1g6r.cd.randions.crd.1 2>&1
cp mdout 1core.serial.log

echo CPU parallel 2,4,8,16 /usr/local/mpich2-1.4.1p1/bin/mpirun
for i in 2 4 8 16 24 32
do
echo $i
mpirun --machinefile=nodefile$i -np $i pmemd.MPI -O -i inp/mini.in -p 1g6r.cd.parm \
 -c 1g6r.cd.randions.crd.1 -ref 1g6r.cd.randions.crd.1 2>&1
cp mdout ${i}core.parallel.log 
done

echo GPU serial
export CUDA_VISIBLE_DEVICES="2"
pmemd.cuda -O -i inp/mini.in -p 1g6r.cd.parm \
 -c 1g6r.cd.randions.crd.1 -ref 1g6r.cd.randions.crd.1 2>&1
cp mdout 1gpu.serial.log

echo GPU parallel 2,4,8,16 /usr/local/mpich2-1.4.1p1/bin/mpirun
export CUDA_VISIBLE_DEVICES="2"
for i in 2
do
echo $i
mpirun --machinefile=nodefile$i -np $i pmemd.cuda.MPI -O -i inp/mini.in -p 1g6r.cd.parm \
 -c 1g6r.cd.randions.crd.1 -ref 1g6r.cd.randions.crd.1 2>&1
cp mdout ${i}gpu.parallel.log 
done

Back