Hi Tat,
It would be best to avoid the older nodes, and run only 1 or 2 MPI jobs
on each newer node, with 12 or 6 threads (-j) each, respectively.
HTH,
S
On 07/16/2014 02:16 PM, Tat Cheung Cheng wrote:
> Hi Dr Scheres,
>
> Sorry to say that I am still a little confused here. Please allow me
> to be a little more specific here.
> I ran refinement on a 12 node cluster, the older half of the 12 nodes
> is with 8 cores each, ~1.5 Gb memory per core, while the newer half is
> with 12 cores each, ~2.4Gb per core. Originally I ran the job with j
> -8, and the estimated memory may exceed the capacity on the older
> nodes, and that may caused the abort.
>
> Estimated memory for expectation step > 8.70496 Gb
>
> Estimated memory for maximization step > 15.3861 Gb
>
>
> I think what I can do here is to use a lower -j, or limit the job to
> the newer nodes with more memory. So my question here is which way do
> you think is a better/more efficient way to run?
>
> Thanks very much again.
>
>
> Best Regards,
>
> Tat
>
>
>
>
> On Tue, Jul 15, 2014 at 3:29 AM, Sjors Scheres
> <[log in to unmask] <mailto:[log in to unmask]>> wrote:
>
> Hi Tat,
> 1: RELION will only include the resolution shells it deems
> necessary in
> the 2D/3D refinement/classification. This is normal.
> 2: You run out of RAM. Are you running a single MPI node and multiple
> threads on each 48Gb node? If you run multiple MPI nodes, each of them
> will replicate the RAM. Threads share the same RAM. See Wiki for more
> details.
> HTH, S
> > Hi Dr Scheres,
> >
> > I have collected some data recently, and I am using Relion for 2D
> > classification, 3D classification and refinement. But there are some
> > questions that puzzle me.
> >
> > 1. In 2D and 3D classification, it seems the signal
> (rlnSsnrMap) drops to
> > zero beyond resolution 8.0A in every class in 2D classification
> and 8.6A
> > in 3D classification, no matter how many particles in the class (an
> > example is shown below), but I have set --strict_highres_exp 6
> in both
> > run. I wonder is it a feature of Relion to use 8.0A/8.6A as a
> cutoff for
> > 2D/3D classification or is there some problem with my original data.
> >
> > 47 0.122396 8.170213 0.001602 0.000000
> > 1.123953e-08 9.000828e-12 2.063652e-05
> > 48 0.125000 8.000000 0.001286 0.000000
> > 1.555068e-08 9.999181e-12 6.504905e-06
> > 49 0.127604 7.836735 0.000000 0.000000
> > 0.000000 0.000000 0.000000
> >
> >
> > 2. In Refinement, the program somehow aborts at the final stage
> when
> > calculating the gold-standard FSC. The error log says
> >
> > Allocate: No space left
> > File: ./src/multidim_array.h line: 962
> >
> > And I increase the max memory (to 48Gb) for the run, but the
> program says
> > Estimated memory for expectation step > 8.64817 Gb, available
> memory =
> > 48 Gb.
> > Estimated memory for maximization step > 15.3861 Gb, available
> memory =
> > 48 Gb
> >
> > I thought 48Gb should be more than enough, but it still aborts
> there. I
> > wonder is the memory I allowed too little for the job? or is there
> > something else I should change.
> >
> > Thanks very much
> >
> > Best Regards,
> > Tat
> >
>
>
> --
> Sjors Scheres
> MRC Laboratory of Molecular Biology
> Francis Crick Avenue, Cambridge Biomedical Campus
> Cambridge CB2 0QH, U.K.
> tel: +44 (0)1223 267061 <tel:%2B44%20%280%291223%20267061>
> http://www2.mrc-lmb.cam.ac.uk/groups/scheres
>
>
--
Sjors Scheres
MRC Laboratory of Molecular Biology
Francis Crick Avenue, Cambridge Biomedical Campus
Cambridge CB2 0QH, U.K.
tel: +44 (0)1223 267061
http://www2.mrc-lmb.cam.ac.uk/groups/scheres
|