On Thu, Apr 1, 2010 at 3:24 PM, Glynn Clements <glynn@gclements.plus.com> wrote:
Jordan Neumeyer wrote:
Just kind of my thought process about how I would try to go about
parallelizing a module.The main issue with parallelising raster input is that the library
keeps a copy of the current row’s data, so that consecutive reads of
the same row (as happen when upsampling) don’t re-read the data.For concurrent access to a single map, you would need to either keep
one row per thread, or abandon caching. Also, you would need to use
pread() rather than fseek()+read().It sounds like you’re talking about parallelism in I/O from a file or
database. Neither of which is my intent or goal for this project. I will
parallelize things after they have already been read into memory, and tasks
are processor intensive. I wouldn’t want parallelize any I/O, but if I were
to optimize I/O. I would make all operations I/O asynchronous, which is can
mimic parallelism in a sense. Queuing up the chunks of data and then
processing them as resources become available.Most GRASS raster modules process data row-by-row, rather than reading
entire maps into memory. Reading maps into memory is frowned upon, as
GRASS is regularly used with maps which are too large to fit into
memory. Where the algorithm cannot operate row-by-row, use of a tile
cache is the next best alternative; see e.g. r.proj.seg (renamed to
r.proj in 7.0).
That makes more sense. So a row is like chunk from the map data? Kind of like the first row of pixels from an image. So from the first pixel to width of image is one row, then width plus one starts the next, and so on and so forth. How large are the rows generally?
Holding an entire map in memory is only considered acceptable if the
algorithm is inherently so slow that processing a gigabyte-sized map
simply wouldn’t be feasible, or the access pattern is such that even a
tile-cache approach isn’t feasible.In general, GRASS should be able to process multi-gigabyte maps even
on 32-bit systems, and work on multi-user systems where a process
cannot assume that it can use a significant proportion of the system’s
total physical memory.
Which is good. I didn’t realize how big the data set could be. What’s biggest map you’ve seen?
It’s more straightfoward to read multiple maps concurrently. In 7.0,
this case should be thread-safe.Alternatively, you could have one thread for reading, one for writing,
and multiple worker threads for the actual processing. However, unless
the processing is complex, I/O will be the bottleneck.I/O is generally a bottleneck anyway. Something always tends to be waiting
on another.When I refer to I/O, I’m referring not just to read() and write(), but
also the (de)compression, conversion and resampling, i.e. everything
performed by the get/put-row functions. For many GRASS modules, this
takes more time than the actual processing.
I can see why, especially for big maps since it’s doing that row-by-row.
So when a GRASS module loads a map the basic algorithm looks something like:
- Read row
- get-row function does necessary preprocessing
- row is cached or held in memory. Does the caching take place after
- row is processed
- Display/write process ? (Or is this after a couple iterations, all of them?)
- repeat (1)
Would it be beneficial/practical to parallelize some of the preprocessing like conversion and resampling before the caching occurs?
Finally, the thread title refers to libraries. Very little processing
occurs in the libraries; most of it is in the individual modules. So
there isn’t much scope for “parallelising” the libraries. The main
issue for library functions is to ensure that they are thread-safe.
Most of the necessary work for the raster library has been done in
7.0.
I was trying to refer to all of the raster modules as a whole, but library is just what the modules share. I’ve changed the title from Parallelization of Raster and Vector libraries to Parallelization of Raster and Vector modules.
Would I be working on GRASS 6.x or 7.x? Is there a minimum compiler version when using GCC/MingW? Just curious because openMP tasks are only supported on GCC >= 4.2. Which may or not be useful, but can be a valuable tool when you don’t know how much data or how many “tasks” you have. Like processing a linked-list or binary trees.
–
Glynn Clements <glynn@gclements.plus.com>
~Jordan
I kid, but that is a very large file… it’s bigger than my current desktop’s hard drive.
