Sunday, December 25, 2016

New fat compute node arrived

New fat compute node arrived. It contains an i7-6800K@3.4GHz, an nvidia GTX-1070, 16 GBytes of physical memory and two WD Red 1TB disks. After a few problems with setting-up the disks as RAID0 (using Intel's rapid storage), a minimal CentOS 7.3 was installed and the nvidia drivers compiled and installed.

First tests with NAMD 2.12 were impressive : on a small peptide problem with 4400 atoms, the new box delivered ~330 nanoseconds per day, a speed-up by a factor of 10 compared with NAMD 2.9 running on a node containing an 8-core AMD FX-8150 plus a GTX 570. To make the comparison hardware-dependent-only, we should test with NAMD 2.9 on the new box (testing new NAMD on the old box won't do because support for fermi GPUs was dropped with NAMD 2.12).

On the same peptide system (4300 atoms) and with a 2.5 fs time step, 470 ns/day were recorded. This reached an outstanding 550 ns/day for a smaller (3300 atoms) system. Think about it again : more than 1 μs in two days. Two days.

Update : It seems that with this hardware and NAMD 2.12 it is necessary to disable off-loading FFT to the GPU by using the 'usePMECUDA no' configuration option.

No comments:

Post a Comment