Issues with the 2.7b3 release of NAMD.

Please see the release notes at http://www.ks.uiuc.edu/Research/namd/2.7b3/notes.html

TclForces (at least the loadcoords command) crashes on Windows due to a bug exposed by Tcl 8.5. Fixed in CVS. The most recent working released Win32 binary is version 2.6, which uses an older version of Tcl.

If you experience hangs at startup while determining CPU topology, try adding +skip_cpu_topology to the command line. If that works then the problem is likely contention at the DNS server. Possible solutions are to enable nscd (network caching daemon) or to add each host's IP address to its own /etc/hosts file and be sure the hosts line in /etc/nsswitch.conf has files as its first entry.

Exits during startup if unable to find parameters for all angles and dihedrals in psf file, even if those angles or dihedrals will be removed by the alchemy module because they contain both reactant and product atoms. Fixed in CVS. Workaround is to add fake parameters for reported atom type combinations to parameter file.

FATAL ERROR: Bad global exclusion count! during minimization stage of "Shape-Based Coarse Graining" tutorial. Pairlist distance is very close to periodic cell size, resulting in double-counting of exclusions when bonds are stretched during minimzation. Fixed in CVS. Workaround is to reduce pairlistdist while optionally increasing margin by the same amount to preserve patch count.

FATAL ERROR: Bad global bond count! immediately following revert command. Fixed in CVS. No workaround.

DrudeBondLen restraint was applied to lone pair bonds. Fixed in CVS. No workaround.

Lone pairs used in Drude force field may be positioned incorrectly (at wrong angles) if the patch grid has only one patch in some dimension (e.g., 3 x 1 x 2) and lone pairs have wrapped around periodic cell. Lone pairs within water molecules (such as TIP4P) are not affected. Fixed in CVS. Workaround is to use twoawayx, twoawayy, or twoawayz to force patch splitting in narrow dimensions.

SMP builds using grid forces hang in first step when running on more than 64 cores. Fixed in CVS. Workaround is to use non-SMP build.

CPU topology determination is wrong for Cray XT SMP builds. The only visible symptom is that the number of unique compute nodes reported at startup is one larger than the correct value (run on 20 nodes, NAMD reports 21). This is a bug in Charm++ 6.2.1 and may degrade parallel scaling performance.

FATAL ERROR: Error on write to binary file ...: Interrupted system call

Ancient bug more common on ibverbs builds, now fixed in CVS. Adding +netpoll to the command line for 2.7b3 (or older) released binaries could (but after testing doesn't appear to) eliminate these errors.

The big technical update with this release is Charm++ 6.2.1, which allows you to start ibverbs jobs with mpirun (even though they don't use mpi). See the release notes and the Charm++ docs (once they are updated).

On Abe at NCSA:

/u/ac/jphillip/NAMD_scripts/runbatch_2.7b3

now uses the ibverbs build.

/u/ac/jphillip/NAMD_scripts/runbatch_2.7b3_smp

uses one process per node for increased maximum available memory, but is slightly slower since one core per node is reserved for the communication thread.

On Ranger at TACC:

/share/home/00288/tg455591/NAMD_scripts/runbatch_27b3

uses ibverbs.

/share/home/00288/tg455591/NAMD_scripts/runbatch_27b3_smp

can use 1way, 2way, or 4way processes per node with 15, 7, or 3 compute threads per process, and is again progressively slower because more cores are used for communication.

On Kraken at NICS:

/nics/b/home/jphillip/NAMD_scripts/runbatch_27b3

still uses MPI but the binary is built with g++ now so it should be noticably faster.

/nics/b/home/jphillip/NAMD_scripts/runbatch_27b3_smp

will use 11 compute threads per process.

The memory optimized build is broken, so you'll want to stick with the builds from April for those projects. As an alternative, the SMP builds share molecular data within a node, allowing larger simulations to run on a given machine.