next up previous contents index
Next: CUDA GPU Acceleration Up: Running NAMD Previous: IBM POWER Clusters   Contents   Index

CPU Affinity

NAMD may run faster on some machines if threads or processes are set to run on (or not run on) specific processor cores (or hardware threads). On Linux this can be done at the process level with the numactl utility, but NAMD provides its own options for assigning threads to cores. This feature is enabled by adding +setcpuaffinity to the namd2 command line, which by itself will cause NAMD (really the underlying Charm++ library) to assign threads/processes round-robin to available cores in the order they are numbered by the operating system. This may not be the fastest configuration if NAMD is running fewer threads than there are cores available and consecutively numbered cores share resources such as memory bandwidth or are hardware threads on the same physical core.

If needed, specific cores for the Charm++ PEs (processing elements) and communication threads (on all SMP builds and on multicore builds when the +commthread option is specified) can be set by adding the +pemap and (if needed) +commap options with lists of core sets in the form ``lower[-upper[:stride[.run]]][,...]''. A single number identifies a particular core. Two numbers separated by a dash identify an inclusive range (lower bound and upper bound). If they are followed by a colon and another number (a stride), that range will be stepped through in increments of the additional number. Within each stride, a dot followed by a run will indicate how many cores to use from that starting point. For example, the sequence 0-8:2,16,20-24 includes cores 0, 2, 4, 6, 8, 16, 20, 21, 22, 23, 24. On a 4-way quad-core system three cores from each socket would be 0-15:4.3 if cores on the same chip are numbered consecutively. There is no need to repeat cores for each node in a run as they are reused in order.

For example, the IBM POWER7 has four hardware threads per core and the first thread can use all of the core's resources if the other threads are idle; threads 0 and 1 split the core if threads 2 and 3 are idle, but if either of threads 2 or 3 are active the core is split four ways. The fastest configuration of 32 threads or processes on a 128-thread 32-core is therefore ``+setcpuaffinity +pemap 0-127:4''. For 64 threads we need cores 0,1,4,5,8,9,... or 0-127:4.2. Running 4 processes with +ppn 31 would be ``+setcpuaffinity +pemap 0-127:32.31 +commap 31-127:32''

For an Altix UV or other machines where the queueing system assigns cores to jobs this information must be obtained with numactl -show and passed to NAMD in order to set thread affinity (which will improve performance):

  namd2 +setcpuaffinity `numactl --show | awk '/^physcpubind/ {printf \
     "+p%d +pemap %d",(NF-1),$2; for(i=3;i<=NF;++i){printf ",%d",$i}}'` ...


next up previous contents index
Next: CUDA GPU Acceleration Up: Running NAMD Previous: IBM POWER Clusters   Contents   Index
http://www.ks.uiuc.edu/Research/namd/