Segfaulting on mpi-linux-amd64 in qt_block

One needs to either edit charm/src/arch/mpi-linux-amd64/conv-mach.h to add #define CMK_THREADS_USE_CONTEXT 1 or use -threads context in the build lines for namd and charm++ test programs.

Brian Bennion

Load balancer hangs on dual-core Opteron clusters, including Cray XD1

There are issues with the Charm++ timer on dual-core Opteron and Athlon 64 systems that can cause the load balancer to hang. A good way to diagnose this is to add "outputtiming 1" and watch for an occasional negative time/step. Clocks running backwards tend to confuse things.

The fix is to edit charm/src/arch/net-linux-amd64/conv-mach.h (or mpi-linux-amd64/conv-mach.h for MPI builds) and change these lines:

#define CMK_TIMER_USE_GETRUSAGE                            0
#define CMK_TIMER_USE_SPECIAL                              0
#define CMK_TIMER_USE_TIMES                                0
#define CMK_TIMER_USE_RDTSC                                1

It works if you move the 1 to CMK_TIMER_USE_TIMES for net-linux-amd64 and for mpi-linux-amd64 you probably want CMK_TIMER_USE_SPECIAL (which uses MPI_Wtime).

This problem is present in 2.6b1 and should be fixed in 2.6b2.

-JimPhillips

I ran a quick apoa1 benchmark for a user trying to choose between Athlon 64 and Opteron:

Athlon 64 3500+ (2.2 GHz)	2.95 s/step
Opteron 148 (2.2 GHz)	2.90 s/step

So, virtually no difference. I've also measured that a dual-core Opteron 100 runs at nearly the same speed as a dual-processor machine with the same clock speed, so I'm guessing that a dual-core Opteron and an Athlon 64 X2 would have indistinguishable NAMD performance.

-JimPhillips

For NAMD 2.5 you need to download the latest Charm++ 5.8 and NAMD from cvs in order to compile for AMD Opteron

Starting from 12/25/2004, charm++ build option "opteron" is obsolete in the cvs version of Charm++, use "net-linux-amd64" or "mpi-linux-amd64":

./build charm++ net-linux-amd64

or,

./build charm++ mpi-linux-amd64

Charm++ 5.8 is untouched.

To build Charm++ for Opteron, use "opteron" option:

./build charm++ net-linux opteron

To build NAMD, use Linux-amd64-g++:

./config ... Linux-amd64-g++

Alternatively, one can compile with MPI, To build Charm++, make sure you have MPI library correctly installed and point "mpiCC" to the correct path. You may want to specify where you installed MPI to build script using option --incdir and --libdir when there is missing <mpi.h> compilation error (see ./build --help for more help)

First build Charm++:

./build charm++ mpi-linux opteron

and build NAMD with Linux-amd64-MPI:

./config ... Linux-amd64-MPI

Build instructions for the namd on an amd64 linux cluster with myrinet (gm2) support.

Notes: This is the process I had to follow to get everything working on our opteron cluster, you may have to customize some of the library and include paths. Also, I used proper installs of tcl and fftw, rather than compiling my own versions. I didn't use the vmd plugins, as apparently our molecule guys don't use it on the actual cluster. :) Damon Smith - Victorian Partnership for Advanced Computing] 14-Jan-2005

In a tmp directory (~/tmp)

* #Retrieve and build Charm++ lib
* cvs -d :pserver:checkout@charm.cs.uiuc.edu:/cvsroot co charm
* cd charm
* ./build charm++ mpi-linux-amd64 gm2 gcc --libdir=/usr/local/mpich-1.2.6-gm-path/lib/ --incdir=/usr/local/mpich-1.2.6-gm-path/include/
* #Namd is a bit confused about charm directories, so we need a symlink here:
* ln -s mpi-linux-amd64-gm2-gcc mpi-linux-amd64
* #Unpack and retrieve namd into ~/tmp/namd
* #Note: you need a recent version of namd to get the mpi-linux-amd64 config.
* cd namd
* vi Make.charm
* #In the file Make.charm, set the CHARMBASE variable as follows:
* CHARMBASE = ${HOME}/tmp/charm
* vi arch/Linux-amd64.tcl
* #In the file Linux-amd64.tcl, set the TCLFLAGS and TCLLIB variables as follows:
* TCLLIB=-ltcl8.4 -ldl
* TCLFLAGS=-DNAMD_TCL -DUSE_NON_CONST
* #And continuing on the command line:
* ./config tcl fftw Linux-amd64-MPI gm2
* cd Linux-amd64-MPI
* vi ~/tmp/charm/mpi-linux-amd64-gm2-gcc/include/conv-mach.sh
* #In this file, conv.mach.sh
* CMK_SYSLIBS="-lmpich -L/usr/gm/lib -lgm"
* make