Re: [AMBER] Clustering getting killed

From: Daniel Roe <daniel.r.roe.gmail.com>
Date: Sat, 22 May 2021 13:29:10 -0400

Forgot to mention you can also strip down the trajectory to only the
atoms you want to cluster, cluster that, write out the cluster number
vs time file, then use that to extract the original trajectory frames.
Let me know if you need help with that,

-Dan

On Sat, May 22, 2021 at 1:28 PM Daniel Roe <daniel.r.roe.gmail.com> wrote:
>
> Hi,
>
> In addition to Carlos's excellent suggestions (use a larger
> sieve/restrict clustering to a smaller subset of atoms) I have a few
> more comments and suggestions
>
> On Sat, May 22, 2021 at 9:43 AM Sruthi Sudhakar
> <sruthisudhakarraji.gmail.com> wrote:
> >
> > trajin gamd2.nc
> > loadcrd gamd2.nc 1 last 5 clus_crd
>
> This is redundant. You're loading the trajectory for processing with
> 'trajin', but then reading that trajectory into a coordinates (COORDS)
> data set with 'loadcrd' named "clus_crd" that you are then not using
> in the subsequent 'cluster' command. So in effect you are using twice
> the memory you need; the coordinates are stored in 'clust_crd' and
> then again in the default coordinates set for "cluster". So either
> remove the 'loadcrd' so that 'cluster' uses the default COORDS set, or
> remove 'trajin' and add 'crdset clus_crd' to your 'cluster' command.
>
> > Estimated memory usage (250000 frames): 83.211 GB
>
> This indicates that you will need at least 83 GB to store the
> coordinates in memory, and then more memory to store the pairwise
> distance matrix, which maybe your analysis node does not have. So my
> suggestion here would be to use the 'loadtraj' command to load the
> coordinates as a 'TRAJ' (i.e. on-disk) data set, e.g.
>
> parm noions.prmtop
> loadtraj gamd2.nc name clus_crd
> cluster crdset clus_crd sieve 10 ...
>
> With 250k frames this means the single-precision pairwise matrix will
> take about 1.2 GB of memory (if I did that math right); cpptraj should
> print an estimate before clustering begins. So if you don't have that
> much memory available you'll want to increase the sieve a bit.
>
> With all that said, I really highly recommend you try the clustering
> on a much smaller subset of your trajectory first (maybe 1000 frames
> or so) just to make sure you have all your settings correct before you
> jump right into clustering 250k frames.
>
> Hope this helps,
>
> -Dan
>
> > ----- gamd2.nc (1-250000, 1) -----
> > 0% 10% Killed
> > _______________________________________________
> > AMBER mailing list
> > AMBER.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber

_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sat May 22 2021 - 10:30:03 PDT
Custom Search