[Developers] EMObject memory management and OD performance

David Epelboim tecnica at geoinfo.com.ar
Mon Jul 23 13:39:34 CEST 2007


Jaap,

Have you considered that if a horizon takes 45% and grows to 50% at the 
time of copying the old to the new array you need to have 95% of the 
size of the horizon allocated in memory?

I know, it's temporary, because after the copy process is ended you'll 
need only 50%, but on those iterative operations you still need memory 
availability close to 100% many times. Plus the times it takes to copy.

I'm not way involved in coding, so I would like to know from you what do 
you think about using dynamic arrays linked to each other using a 
pointer at the last position. Or using several 2D arrays and "tiling"

Regards,

David.


Jaap Glas wrote:
> Dear all,
>
> For those still wondering why operations on e.g. horizons can sometimes
> become really slow, and everyone interested in the internals of OpendTect.
>
>
> PERFORMANCE OF OPENDTECT IN RELATION TO ITS EMOBJECT MEMORY MANAGEMENT
>
> The usual way to store an EMObject that is not too sparsely distributed
> in geometrical space is keeping it in one or more n-dimensional arrays
> in memory. This guarantees quick random access to individual geometrical
> positions. For example, a 3D horizon will be stored in a 2-dimensional
> array, or in several if it consists of more than one section.
>
> The strategy of OpendTect is to keep the sizes of those arrays as small
> as possible to restrict memory usage. We will not reserve an array that
> covers the whole survey if the EMObject is currently located in a small
> part only. The disadvantage of this strategy is of course that a growing
> object does not fit its array any more. What will happen if the function
> EMObject::setPos(,,) likes to add a new position outside the area covered
> by the current array? A bigger array will be allocated and the contents
> of the old array have to be copied to the new array before the old one
> can be released. This takes time, and its duration is of course linearly
> proportional to the current size of the object. However, not wasting
> valuable memory makes it worth the effort.
>
> Unfortunately, this economical memory management strategy has one pitfall.
> A gradually growing EMObject will induce a repeated copying of the object
> to a new and slightly bigger array. The total computational effort becomes
> quadratically proportional to the final size of the object. This can become
> a real burden in case of large EMObjects. Moreover, it is unnecessary if
> either the final size or a reasonable maximum size of the object is already
> known in advance. When applying the autotracker to extend a 3D horizon for
> example, its new maximum size will be bounded by the space covered by the
> tracker box. To transfer such knowledge to the ignorant lower parts of
> OpendTect, the programmer can use the following functions:
>
> 1) void expandWithUdf( RCol& start, RCol& stop ) from Geometry/binidsurface.h
>
>    Reserves memory for a predefined space around the current object in one
>    go, and fills the added space with Undefs. At the moment, such a function
>    only exists for the class BinIDSurface, in which the different sections
>    of e.g. a 3D horizon are stored.
>
> 2) virtual void trimUndefParts() from Geometry/geomelement.h.
>
>    Frees the memory occupied by all undefined space around the current
>    object. To be applied if it is not guaranteed that all reserved space
>    has been used by the operation extending the object. For example, the
>    autotracker will not always succeed in tracking up to the boundaries
>    of the tracker box.
>
> A clear example of the application of the two functions above can be found
> in Horizon3D::setArray2D(,,) in EarthModel/emhorizon3d.cc, which is called
> to provide an existing 3D horizon with new depth values. The performance of
> an operation like "Fill Holes" with option "Extrapolate outward" set does
> really benefit from it. Another example of their application is hidden in
> the autotracker code in MPEEngine/autotracker.cc. Beware that the call to
> expandWithUdf(,) is executed by function Horizon3DExtender::preallocExtArea()
> in this case.
>
>
> Best regards.
>
>   





More information about the Developers mailing list