Here is an update on namd for the shared memory one-node version namd2/namd3
and the multi-node version charmrun++
. The standard NAMD benchmark apoa1 is too small to show the scaling on a reasonably modern system, so here we use a user's lipid simulation for 25k steps until it prints its “benchmark” performance.
Most of NAMD 3 and a few of the newer NAMD 2 are not usable with Centos 7 OS because of being compiled with too new a glibc. The exceptions we have found are 2.15alpha1 (CPU AVX512) and 3.0-alpha7-GPU, which are the best performing runnable versions we have found, but both only have “multicore” or single node shared memory namd2/3
versions. For charmrun++
, a verbs-smp edition such as 2.14 is indicated. Newer versions will be available after we reimage the cluster in Rocky 8 OS.
We are using the number of cores available on the node, either “+p32” or “+p64”, except: NAMD recommends using one fewer core to run than present in the hardware; we find that to be beneficial for Intels and not beneficial for AMDs as reflected in the examples.
The 2.14 verbs-smp
version can be used with both namd2
and charmrun++
. Pinnacle I is over twice as fast as Trestles on this version, and Pinnacle II is over twice as fast as Pinnacle I.
module load namd/2.14 #Pinnacle I Intel 6130 2.09 ns/day namd2 +p31 +setcpuaffinity +isomalloc_sync step7.2_production_colvar.inp #Pinnacle II AMD 7543 0.81 ns/day namd2 +p64 +setcpuaffinity +isomalloc_sync step7.2_production_colvar.inp #Trestles AMD 4.51 ns/day namd2 +p32 +setcpuaffinity +isomalloc_sync step7.2_production_colvar.inp
The 2.15a1 AVX512 version with namd2
here runs only on Pinnacle I, but is very much faster for that case than is 2.14.
module load namd/2.15a1 #Pinnacle I Intel 6130 1.24 ns/day namd2 +p31 +setcpuaffinity +isomalloc_sync step7.2_production_colvar.inp
Single node 2.14 charmrun++ ++np 1
with ++ppn ##
moved to left side should run equivalently to the same namd
and same ++ppn ##
.
With two nodes, in a few cases charmrun++
scales fairly well, but because of better alternatives, the prospects for worthwhile charmrun++
runs are few with this set of compute nodes.
On Pinnacle I, 2.14 charmrun++ ++np 2
scaled well but was still hardly faster than single-node 2.15a1 namd2
. Three nodes didn't scale well at all, so there's not really a good use case for charmrun++
.
module load namd/2.14 #Pinnacle I Intel 6130 1 node 2.09 ns/day charmrun ++remote-shell ssh ++np 1 ++ppn 31 `which namd2`+setcpuaffinity +isomalloc_sync step7.2_production_colvar.inp #Pinnacle I Intel 6130 2 node 1.17 ns/day charmrun ++remote-shell ssh ++np 1 ++ppn 31 `which namd2`+setcpuaffinity +isomalloc_sync #Pinnacle I 3 node 0.88 ns/day charmrun ++remote-shell ssh ++np 1 ++ppn 31 `which namd2`+setcpuaffinity +isomalloc_sync
On Pinnacle II, two-node charmrun++
didn't scale well, so again little use case for charmrun++
.
module load namd/2.14 #Pinnacle II AMD 7543 2 node 0.69 ns/day charmrun ++remote-shell ssh ++np 2 ++ppn 64 `which namd2` +setcpuaffinity +isomalloc_sync step7.24_production.inp
On Trestles, two-node 2.14 scaled well to about the same speed as one-node 2.14 namd2
on Pinnacle I. Three nodes did not scale well. So here there may be a use case for using an uncrowded cluster.
module load namd/2.14 #Trestles AMD 2 node 2.81 ns/day charmrun ++remote-shell ssh ++np 2 ++ppn 64 `which namd2` +setcpuaffinity +isomalloc_sync step7.24_production.inp #Trestles AMD 2 node 1.99 ns/day charmrun ++remote-shell ssh ++np 2 ++ppn 64 `which namd2` +setcpuaffinity +isomalloc_sync step7.24_production.inp
=nodelist=
charmrun++
is expecting, instead of an mpirun
hostfile/machinefile as generated by slurm, a file called nodelist
that resembles this:
host tres0931 host tres0929 host tres0928
To modify the machinefile (generated by the system for each job) to be a nodelist in the PWD, try
cat /scratch/${SLURM_JOB_ID}/machinefile_${SLURM_JOB_ID} | sed "s/^/host /" >nodelist
Overall there is not a good use case for charmrun++
because there are better alternatives, except for utilizing the underused capacity of Trestles.
Here using the number of CPU cores available on the node (24/32/64) and one GPU (two or more GPUs devices 0,1,2,3
scale poorly, not recommended or approved for AHPCC public use partitions).
This is a simulation that takes about 3.5 GB of GPU memory. With a smaller simulation that takes about 400 MB of GPU memory, the CPUs used did not scale and were best around +p4
.
On the gpu72
nodes with Intel 6130 and single NVidia V100, it's about 5 times faster than the best CPU version, so are a good use case. On agpu72
nodes with AMD7543 and single A100, it's only about 10% faster than 6130/V100, so that's not a good use case for the more expensive AMD/A100 nodes, unless gpu memory requires the newer GPU. The even more expensive multi-gpu qgpu72
nodes also don't scale well over single-gpu and are not a good use case.
#gpu72/v100: module load namd/3.0a7 namd3 +p32 +setcpuaffinity +isomalloc_sync +devices 0 step7.2_production_colvar.inp Info: Benchmark time: 32 CPUs 0.0393942 s/step 0.227976 days/ns 0 MB memory #agpu72/a100: not recommended unless memory requires namd3 +p64 +setcpuaffinity +isomalloc_sync _devices 0 step7.24_production.inp Info: Benchmark time: 64 CPUs 0.0344332 s/step 0.199266 days/ns 0 MB memory