Hi,
Some people reported this (RELION takes very long
between iterations; see GitHub issues) but we could
not reproduce it so far.
> (even writing to scratch disk)
Indeed, this was sometimes solved by avoiding 'Preread
particles to RAM' and using scratch instead, but
not always (as you noticed).
> 9674 particles with 910-pixel boxes
Why you need such a large box for Class2D?
4 A/px is enough for Class2D.
> We’ve tried 48 cpus with 1 thread
In general this is a bad idea. If the machine has 48
cores in total, I would use 5 MPI x 12 threads
(MPI rank 0 does nothing) or 9 x 6 for relion_refine.
Best regards,
Takanori Nakane
On 2020/02/19 21:55, David Michael Belnap wrote:
> Takanori,
>
> I have a follow-up question. We’ve been trying to do 2D classification on a data set of 9674 particles with 910-pixel boxes on machines with 128 Gb memory. According to the formula in the RELION “?” help, those particles should take 30 Gb of memory. However, whatever we seem to try (even writing to scratch disk), the process just hangs at the first iteration. Even particles half that size are now hanging. Do you have any suggestions on different settings we could try? We’ve tried 48 cpus with 1 thread (on two nodes giving 256 Gb of memory total). Would more threads help?
>
> Thank you,
>
> David
>
>
>> On Feb 18, 2020, at 4:28 AM, Takanori Nakane <[log in to unmask]> wrote:
>>
>> Hi,
>>
>> Please look at
>> https://www3.mrc-lmb.cam.ac.uk/relion/index.php/Benchmarks_%26_computer_hardware
>> for recommendations on accelerated CPU version.
>>
>>> 1500 pixel box size
>>
>> Just to make sure: do you need this box size even at a reasonable pixel size?
>> Some people "unbin" back to the original detector pixel size
>> even when their resolution is far from the current Nyquist frequency.
>> There is no point doing so.
>>
>> Another consideration is that for an extremely large virus,
>> icosahedral symmetry is often broken and you have to work on
>> smaller sub-volumes. So you might end up not needing such a large box.
>>
>> Best regards,
>>
>> Takanori Nakane
>>
>> ########################################################################
>>
>> To unsubscribe from the CCPEM list, click the following link:
>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
>
>
> ########################################################################
>
> To unsubscribe from the CCPEM list, click the following link:
> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
>
########################################################################
To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
|