taskGNU Astronomy Utilities - Tasks: task #13900, Scalable mesh structure for...

 
 

You are not allowed to post comments on this tracker with your current authentication level.

task #13900: Scalable mesh structure for complex operations

Submitter:  Mohammad Akhlaghi <makhlaghi>
Submitted:  Fri 19 Feb 2016 07:43:06 AM UTC
 
Should Start On:  Thu 18 Feb 2016 03:00:00 PM UTC Should be Finished on:  Thu 18 Feb 2016 03:00:00 PM UTC
Category:  All Gnuastro Priority:  9 - Immediate
Item Group:  Enhancement Status:  Postponed
Privacy:  Public Percent Complete:  0%
Assigned to:  makhlaghi Open/Closed:  Open
Effort:  0.00

Fri 19 Feb 2016 07:43:06 AM UTC, original submission:  

When dealing with large datasets (images or 3D datacubes), it is very memory and time consuming to read the whole dataset into memory before operations. Nearly every time consuming operation in Gnuastro (not ConvertType for example) can be done on a small part of the input data independently. Fortunately CFITSIO can also read and write from and to parts of a file. So I would like to suggest that we design a highly scalabe input/output+operation system, where for example huge datasets (much larger than the RAM) can be easily processed: for example in warping or to run NoiseChisel on.

Fortunately a very rudimentary first step is already in place for the operations of NoiseChisel and SubtractSky: mesh.[ch]. But all these functions are built for use within RAM. Here I am suggesting to extend the definition and implementation in such a way that it merges with operations in fitsarrayvv.[ch] so when a program thread (CPU or GPU, see task #13787) needs only one part of an image, it only reads that one part, does its operations and then saves its results to the proper part of an output file. This way the RAM will never get full and parallel (CPU or GPU) programming can create a huge speed boost.

The proposed mesh structure should be able to allow a certain amount of margin (or extra pixels) for each mesh, so operations like spatial convolution or warping can also be done in this embarrassingly parallel manner and without relying on a huge RAM. Particularly with SSD devices that are becomming popular, the read/write speed is not a major issue any more.

Like NoiseChisel and SubtractSky, the user can set the mesh parameters, so if they don't have any RAM issue, they can simply only define one mesh and they won't have to worry about read/write speeds slowing them down. However, when the datasets are huge (like a MUSE datacube, or one HSC or LSST exposure), this is very important.

If it becomes too complex, this mesh structure can be made over an image in a separate utility and all the mesh parameters can be added as headers in the FITS file. Then any program that needs them will read those headers and use them as a basis to operate on the image in a highly scalable way.

An array can also be defined to keep one value per mesh in operations like finding the Sky value and its standard deviation in NoiseChisel, SubtractSky and MakeCatalog.

These are all just raw suggestions and are a more generalized case for task #13759 (reducing the memory usage of MakeCatalog). Certainly as the implementaion starts, the details can/will change from what is suggested here. The main goal however is to allow a high level of scalability (more parallel CPU and GPU operations and reducing the dependancy on the RAM)

Mohammad Akhlaghi <makhlaghi>
Project Administrator

 

(Note: upload size limit is set to 16384 kB, after insertion of the required escape characters.)

No files currently attached

 

Depends on the following items: None found

Items that depend on this one: None found

 

Carbon-Copy List
  • -email is unavailable- added by makhlaghi (Submitted the item)
  •  

    There are 0 votes so far. Votes easily highlight which items people would like to see resolved in priority, independently of the priority of the item set by tracker managers.

     

     

     

    Follows 1 latest change.

    Date Changed by Updated Field Previous Value => Replaced by
    2016-02-19 makhlaghi SummaryScalable mesh structure for input/output Scalable mesh structure for complex operations

    Back to the top


    Powered by Savane 3.9