Re: [Yt-dev] Geometry, RAMSES and non-patch datasets

Hi Sam, I'm glad you're interested in helping to develop yt's Ramses facilities. yt is a community code, and we're always excited to embiggen that community. However, I think you might want to take a step back and look at what yt already is and does. yt is highly optimized already for patch based AMR. We have thought through these concepts in significant detail, and over the past nearly 5 years we have refactored and optimized the loading of data many times. I don't know if this is your intention, but you seem to be suggesting that we have not thought about iterators and where the bottlenecks lie in our analysis to data pipeline. We did not find it necessary to jump to C often (we do not use C++ because of portability issues on some of the supercomputing centers we deploy on). Your emails seem to suggest you believe a rewrite of many core routines in C/C++ would be necessary. I think that the Ramses reader could be brought to speeds comparable to the enzo/orion readers by optimizing the approach already pursued by those yt frontends. Python has extremely powerful object orientation; the need for C++ templates is obviated by those features, and we have already demonstrated that a small amount of C can go a very long way for speed. Again, I want to reiterate that we are excited that you are interested in contributing; I myself am working with a group using Ramses, and it would be great to have an optimized frontend. I just want to impress upon you that there is a lot of people in the yt community, and we have indeed thought many of these issues through. It might be helpful to take a look at how the code works (there is a sample enzo dataset in the distribution), and feel free to ask questions on the dev list, on the user list, and on #yt, if you're into the whole IRC thing. sincerely, Jeff Oishi On Wed, Jun 22, 2011 at 8:02 AM, Sam Geen wrote:

So I guess a pattern for stepping through the iterators could go: - Step through the iterators in the normal Python way. - The next function checks to see whether we require extra resolution (somehow...) - If we do, it jumps to its children, if not it jumps to its neighbours. - Whether or not it needs to open a new file/grid is determined under-the-hood. If it's too slow to do this on a per-iterator basis, the overarching container class (whether that's the snapshot or the AMR grid or whatever) could jump through files/grids, which then give iterators to their own grid cells which follow the above pattern. I guess if this is slow then it could always be folded into C++, but I'd need to look more into Cython to see how easy it is to do, since the analysis algorithms will probably need to talk to the iterators. Like you say, numpy is quite fast at doing array operations, and I don't know if the iterator pattern fits into that, since iterators tend to operate individually. Maybe the C++ I/O could do this iterator pattern and then offer up arrays in memory to be operated on. Alternatively, the analysis algorithms could be chunked up into C++ routines, but I guess this makes it harder to write analysis routines. Then again, no reason YT can't have both, with analysis routines being refactored into C++ too. In general, iterators have been pretty successful as a way of coupling diverse container classes to diverse algorithms in C++, although I don't know how their performance scales in Python.

I'm happy to work in tandem on this. My timetable is that I'm moving into a new job around the end of July, so I might be a bit up-in-the-air during August, but can probably make time to look into this around then. I'm happy to do some C++ work, but my guess is that it's better to start with everything in Python and then figure out where the bottlenecks are with the profiling tools.

Sam

On 22/06/2011 15:44, Matthew Turk wrote:

...

Hi Sam,

On Wed, Jun 22, 2011 at 6:11 AM, Sam Geen  wrote:

...

OK, cool. So is there anything that I can do to help with the Ramses instantiation? I could just play about with implemeting Pymses with the current I/O interface, but if you're already doing this then I'll wait to see what you come up with. Also, if the I/O interface is changing soon then I might hold off.

There's definitely stuff you can do to help out; if you want to give a shot at profiling we can figure out some of the places it might still be struggling.  For instance, if you have a file that just includes:

import cProfile pf = load("output_00007/info_00007.txt") cProfile.run("pf.h.print_stats()", "output.cprof")

it should create an output.cprof that you can visualize either using the pstats module or the pyprof2html ("pip install pyprof2html", although you may have to also pip install jinja) to see where the bottlenecks are.  My guess is that calculating the hilbert indices is still a big portion, as that's what it was for me.

...

Actually, if there's any design documentation not in the code paper or not obviously on the website floating around then it might be useful to have a look at this. Also, another vaguely related query - how much can Python talk to compiled code (C++, etc)? To what extent can you pass Python objects to C++ routines, assuming you know what the public methods are?

Right now unfortunately the only design documentation is either in the paper, or ad hoc in the docs.  I'm happy to work through some of that here, though, and it might be a good idea to consolidate this information on the wiki, too.

As for C++, it's actually very straightforward; we are using Cython (not Boost::Python, which I am told is quite nice but unfortunately carries with it the Boost overhead) and you can see how this is done primarily in yt/frontends/ramses/_ramses_reader.pyx, which uses headers found in yt/frontends/ramses/_ramses_headers , courtesy of Oliver Hahn.

The more I think about it, the more viable I think it is to move all the grid selection routines into the geometry class (although I am not committing myself to moving fundamental routines *outside* of the objects they relate to, just yet) and then using them as the fluid blob selection routines.  I have to think about how to do this without relying on bounding box arguments, but it should be viable.  I would like to work on this together -- perhaps after I am back from vacation we can try to start refactoring the necessary classes?  The testing infrastructure is coming into place rapidly, so it would be feasible to try to do this.

Thanks for your thoughts on this.

-Matt

...

Sam

Matthew Turk wrote:

...

Hi Sam,

On Tue, Jun 21, 2011 at 8:07 AM, Sam Geen  wrote:

...

Hi,

This is in reply to Matt's e-mail from 3 weeks ago (I only just realised I forgot to hit "confirm" on the yt-dev mailing list signup).

No worries!  :)

...

I guess one solution to the problem would be to abstract what a "grid" is (I'm guessing a grid is a container for a geometrically consistent chunk of the entire simulation volume?) Then allow it to answer queries about its geometric properties itself. So for example, ask it "myGrid.IsInRegion(myWeirdGeometricConstruct)". I guess the trick is to figure out a flexible but simple interface for this, depending on how well you know the requirements for what the grid should be able to do. In general, I think this is the ideal situation, because as Matt says hammering every code into the same structure in memory creates slowdowns. One possibility is to create a few template memory structures, etc, to allow people to bolt together new implementations for each code.

A grid is indeed a container for a chunk of the simulation -- typically in patch-based AMR codes, these will be some (hopefully large but not too large) contiguous region.  This enables numpy to take over, as it helps batch mathematical operations -- for instance, for an operation like:

field1*field2

the startup cost of parsing, identifying the object as a buffer of contiguous data, identifying the types, dispatching the correct function, and then allocating and returning a new buffer is the startup cost against which the actual operation of multiplication is weighed.  The batching of operations with grids nicely coincides with reducing the ratio between startup to operation cost.

Right now, the mechanism for geometric constructs is inverted from how you describe -- when describing a sphere, for instance, the operation is:

 * Query the Hierarchy object (which I would support renaming to 'mesh' or 'geometry' in a future iteration of the code, likely 3.0) to identify grids that intersect the geometric region.  This is accomplished through a "geometry mixin" that supplies various routines to do this.  * Query each intersecting grid's x,y,z values (for each cell) to identify which cells intersect the region.  * Return these values to the routine/user that requested them.

I think that this is compatible with what you have outlined, in general.  The issue I had hoped to avoid was to reduce the interaction between IO and geometry as much as possible, simply because IO routines are usually compiled code, whereas ideally I would like the geometry to be performed in Python.  (As it stands it's usually done with operations on bounding values of grids to find intersections.)

...

In terms of choosing algorithms for different types of fluid blob (e.g. one for particles, one for grids), this can be done using functionoids for the algorithms (or at least functionoid wrappers) and then a functionoid factory for spawning the correct functionoid to use with the container. You'd have to wrap all this up in a simple interface again, otherwise it'd be impossible to use.

I also suggested to Matt to create a "fluid blob" iterator that works for all types of fluid blob (SPH particle, octree grid cell, voronoi tessellation cell) but this might be very slow in Python. That said, iterating over "grid"s as chunks of the amr grid instead is a possibility. Having some kind of iterator option might be good, though, as doing things like tracking particles through different snapshots is something I've been doing extensively in my (pre-YT) work.

A generalized fluid blob iterator would be interesting; I think into this the grids could be placed.  By extending the geometry mixin to work with different methods, this could be feasible.  I wonder if perhaps rethinking the idea of static geometries (determined at instantiation) would assist with addressing SPH data.  I am inclined to think this would be a way forward.  In looking over the code, it's not clear to me that there are many places that grids are assumed except in the projections and the first-pass of data selection.

(Projections as we do them now might port nicely to SPH, but it's not yet clear to me.)

...

I don't know how much of this is already known; my domain is Ramses, which is still very slow to use with my dataset (although Matthew has been very helpful in working on the Ramses side of things). I thus haven't looked too much at YT yet as it's still prohibitively slow to load my dataset and play with it.

I did manage to squeeze out what for me was an OOM improvement in RAMSES data instantiation, but I confess it is still slow.  And there are other issues with it.  Right now Casey and I are refactoring fields, and I have set up a testing infrastructure, so I am feeling bit more inclined to try more invasive changes after branching into 2.3 and 3.0 branches sometime later this summer.

Perhaps moving to a generalized geometry, into which the standard patch/block AMR "hierarchy" paradigm would fit, would meet the necessary needs to do generalized fluid operations...

-Matt

...

Cheers,

Sam

On Tue, Jun 7, 2011 at 16:15 AM, Matthew Turk wrote:

Hi all,

This is a portion of a conversation Sam Geen and I had off-list about where to make changes and how to insert abstractions to allow for generalized geometric reading of data; this would be useful for octree codes, particles codes, and non-rectilinear geometry.  We decided to "replay" the conversation on the mailing list to allow people to contribute their ideas and thoughts.  I spent a bit of time last night looking at the geometry usage in yt.

Right now I see a few places this will need to be fixed:

 * Data sources operate on the idea that grids act as a pre-selection for cells.  If we get the creation of grids -- without including any cell data inside them -- to be fast enough, this will not necessarily need to be changed.  (i.e., apply a 'regridding' step of empty grids.)  However, failing that, this will need to be abstracted into geometric selection.  For cylindrical coordinates this will need to be abstracted anyway.  The idea is that once you know which grids you want, you read them from disk, and then mask out the points that are not necessary.  * The IO is currently set up -- in parallel -- to read in chunks. Usually in parallel patch-based simulations, multiple grid patches are stored in a single file on disk.  So, these get chunked in IO to avoid too many fopen/seek/fclose operations (and the analogues in hdf5.) This will need to be rethought.  Obviously, there are still some analogues; however, it's not clear how -- without the actual re-gridding operation -- to keep the geometry selection and the IO separate.  I would prefer to try to do this as much as possible.  I think it's do-able, but I don't yet have a good strategy for it.

My current feeling now is that the re-gridding may be a slightly necessary evil *at the moment*, but only for guiding the point selection.  It's currently been re-written to be based on hilbert curve locating, so each grid has a unique index in L-8 or something space.

I believe that geometry and chunking of IO are the only issues at this time.  One possibility would actually be to move away from the idea of grids and instead of 'hilbert chunks'.  So these would be the items that would be selected, read from disk, and mapped.  This might fit nicer with the Ramses method.

What do you think?

Best,

Matt

_______________________________________________ Yt-dev mailing list Yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org

_______________________________________________ Yt-dev mailing list Yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org

_______________________________________________ Yt-dev mailing list Yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org

_______________________________________________ Yt-dev mailing list Yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org

_______________________________________________ Yt-dev mailing list Yt-dev@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-dev-spacepope.org