Hi Eric and Sjors,
It appears the same happening to the particles of 100-200 KDa we are working on. Binning by 2 is what we are testing. I wonder if band-pass filtering or phase-flipping before particle extraction could be helpful in this case.
We also noticed that in our HPC, using 27 CPUs and 4 K80 GPUs was the fastest combination in processing. We assigned 3.8GB RAM per CPU and 3 CPUs per task. But once a while we had segmentation fault and memory mapping errors in a thread after the 10-20 runs; we then could not continue from the exiting points due to recurring segmentation faults. Our sys admin thinks it is memory management issue. I hope that binning will solve the issue as a first step. Did anyone encounter a similar issue before? A costly solution is to assign more Threads per task as our RAM assignment goes with NPU. I wonder if we can do it differently¡A say using a program to control RAM usage for master and slaves. Thanks for sharing.
Best,
Q-X Jiang
MCS @ UF IFAS | EM @ UF ICBR
Sent from iPhone. Pls excuse typos if any.
> On Jan 7, 2018, at 5:00 PM, Eric Hanssen <[log in to unmask]> wrote:
>
> HI Sjors,
>
> Thank you. I guess it is likely since the particles are much smaller, but they are form the same micrographs then the bigger particles I was working with before (killing two birds with one stone may not be such a good idea ...)
>
> Regards
> Eric
>
>
> -----------------
> Assoc. Prof Eric Hanssen
>
> Head - Advanced Microscopy Facility
> Honorary Principle Fellow ¡V Department of Biochemistry and Molecular Biology
>
> Bio21 Molecular Science and Biotechnology Institute
> 30 Flemington Road - The University of Melbourne - Victoria 3010 - Australia
> email: [log in to unmask] | Office: +61 3 83442449 | Microscope: +61 3 83442509
> Web: www.microscopy.unimelb.edu.au
>
>
>
>
> -----Original Message-----
> From: Sjors Scheres [mailto:[log in to unmask]]
> Sent: Friday, 5 January 2018 6:23 PM
> To: Eric Hanssen
> Cc: [log in to unmask]
> Subject: Re: [ccpem] 3D classification speed
>
> Hi Eric,
> Perhaps the SNR is much lower? That would slow down calculations too, as the probability distributions get much broader.
> HTH,
> Sjors
>
>> Hi all,
>> I have speed issue that I don¡¦t understand on my quad GTX 1080 system.
>> I have been working with biggish particles, 340 pixel box size, and
>> lowish number 22,000 particles and each iteration of the 3D
>> classification was about 3min on average. I have now moved to
>> something much smaller, 110 pixel box size and 65,000 particles but
>> now the average 3D classification iteration is a little over 25min.
>> All the other options are the same but obviously for the initial model.
>> Is there a computational explanation or do I need to dig and see if my
>> hardware has issues?
>>
>> Assoc. Prof Eric Hanssen
>> Head - Advanced Microscopy Facility
>> Honorary Principal Fellow ¡V Department of Biochemistry and Molecular
>> Biology
>>
>> Bio21 Molecular Science and Biotechnology Institute - 30 Flemington
>> Road - The University of Melbourne - Victoria 3010 - Australia
>> email: [log in to unmask]<mailto:[log in to unmask]> | Office:
>> +61 3 83442449 | Microscope: +61 3 83442509
>> Web:
>> www.microscopy.unimelb.edu.au<https://urldefense.proofpoint.com/v2/url?u=http-3A__www.microscopy.unimelb.edu.au_&d=DwIGaQ&c=pZJPUDQ3SB9JplYbifm4nt2lEVG5pWx2KikqINpWlZM&r=P2JY4Ggp34jCrKsRhFNcaQ&m=7TlA92lyJJuyuWZTSS0JgNPTAzl7NatUCHuyHuuiQCk&s=7V2BEP_348edYGRXP7uxjz3chDAwQ9hebWh1JPapGQ4&e= >
>>
>>
>
>
> --
> Sjors Scheres
> MRC Laboratory of Molecular Biology
> Francis Crick Avenue, Cambridge Biomedical Campus Cambridge CB2 0QH, U.K.
> tel: +44 (0)1223 267061
> https://urldefense.proofpoint.com/v2/url?u=http-3A__www2.mrc-2Dlmb.cam.ac.uk_groups_scheres&d=DwIGaQ&c=pZJPUDQ3SB9JplYbifm4nt2lEVG5pWx2KikqINpWlZM&r=P2JY4Ggp34jCrKsRhFNcaQ&m=7TlA92lyJJuyuWZTSS0JgNPTAzl7NatUCHuyHuuiQCk&s=8IpFpUFSPSs2O_axljUv_cjJBdD0uB8FiUPuoz4OwOI&e=
>
|