Hello,
I am simulating a periodic crystalline system with GPU accelerated lj96/cut and coul/long pair styles using the most recent stable version of LAMMPS (1 Feb 2014). My input script runs fine on an older Tesla card with the following deviceQuery output:
Device 0: "Tesla C2070"
CUDA Driver Version / Runtime Version 6.0 / 5.0
CUDA Capability Major/Minor version number: 2.0
Total amount of global memory: 5375 MBytes (5636554752 bytes)
(14) Multiprocessors x ( 32) CUDA Cores/MP: 448 CUDA Cores...
On this card, I compiled the GPU library with double/double precision and the -arch=sm_21 flag using cuda/5.0.35.
On a newer Tesla K20Xm card, I compiled the GPU library with -arch=sm_30 using both cuda/5.0.35 and cuda/6.0.37. In either case, the pressure blows up and the simulation crashes after 1 step on the K20Xm if a long range Coulombic solver is used with lj96/cut/gpu. I've tried these combinations of pair styles using "hybrid/overlay":
lj96/cut/gpu, coul/long/gpu, pppm/gpu -> blows up
lj96/cut/gpu, coul/cut, none -> works fine
lj96/cut, coul/long/gpu, pppm/gpu -> works fine
lj96/cut/gpu, coul/long/gpu, pppm -> blows up
lj96/cut/gpu, coul/long/gpu, ewald -> blows up
lj/cut/gpu (12-6 style), coul/long/gpu, pppm/gpu -> works fine
The problem seems to be isolated to when lj96/cut/gpu is used in conjunction with a long range solver.
The deviceQuery output for this card is:
Device 0: "Tesla K20Xm"
CUDA Driver Version / Runtime Version 6.0 / 5.0
CUDA Capability Major/Minor version number: 3.5
Total amount of global memory: 5760 MBytes (6039339008 bytes)
(14) Multiprocessors x (192) CUDA Cores/MP: 2688 CUDA Cores
To test if this problem is isolated to my local cluster, I tested my code on Amazon's EC2 servers. On a cg1.4xlarge instance running a Tesla C2070, everything works fine. On a g2.2xlarge instance running a newer compute capabaility 3.5 GPU card, I get the same pressure explosion problem.
My input and data files are too large to post here, but I can email them you if you'd like to try them. I'd really appreciate any feedback.
Thanks,
Jeff Camp
Sholl Group
Georgia Tech