On Tue, Jan 27, 2015, Rosemary Mantell wrote:
>
> I should probably mention that I first saw this problem when using the
> AMBER 12 DPDP model, so it's not just a problem with fixed precision.
Let me try to clarify things a bit here:
1. We do not expect the DPDP model to give exactly the same answer every time.
This arises because floating point calculations depend on the order of
computation, and this will change from one run to the next.
2. However, run-to-run differences in DPDP should be very small for a
single energy calculation. This was why I asked earlier about how big
the differences were with DPDP. If the run-to-run diffs with DPDP are as
large as you report for DPFP, something must be amiss. If you get diffs
that are more like round-off errors, that is more in line with what we
expect.
3. SPFP and DPFP should be reproducible, since all parallel reductions
are done with fixed point arithmetic, which is independent of the order
of calculation. The fact that you see reproducibility with SPFP but not
DPFP is pointing the finger at the latter. This is all the more true
since very few people ever use the DPFP mode, and there might be something
lurking there that we haven't come across before.
4. One possible explanation for the DPFP failure is that you have unusual
dynamic range in your exmaple, such that the fixed point representation is
overflowing. This is why people are eager to know what the RMS gradient
of your structure is, since "bad" initial structures could have large
terms that violated the fixed-point assumptions. (This is the reason for
the general recommendation that initial minimizations, which are very
cheap, be carried out on a CPU.) Although you haven't explicitly said so,
it sounds like this is not the case here.
5. It should be up to us to track this down, and I know Ross is working
with your input files. But if you come across useful information (such
as: the problem goes away if you turn of gb; or set igb to some other
value; or do the calculation on a smaller system, etc.) please let us
know.
...thx...dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Jan 27 2015 - 12:30:03 PST
