follow

#348 Quad proj doesn't scale

Reported by Matthew Turk 2011-12-06

The quadtree projection does not scale particularly well (although it performs quite well in serial.) The merging algorithm requires transforming buffers into data structures, merging, and then transforming back. We should move to a system with a single transform, stratified by level with a PH key for the value, and stepping along it in 1D in sync between two trees to add them up.

This will require no modifications to the Python code for the projection itself, but the frombuffer/tobuffer code will require modification, as will the merging code.

One potential problem is that with 64-bit integers we may run out of positions in the PH curve for some types of simulations (i.e., mine.)

Attachments

Comments and changes

1. #1 Matthew Turk

   written 2012-04-30

   • Changed status from new to open.

   Sam has an idea: simply make the tree for all the grids on all the processors, but don't fill them except on the processor that owns the grid. I think this might work.
2. #2 Matthew Turk

   written 2012-05-03

   First pass: https://bitbucket.org/MatthewTurk/yt/changeset/3f39eb7bf468
3. #3 Matthew Turk

   written 2012-05-03

   • Changed responsible from nobody to MatthewTurk.
4. #4 Matthew Turk

   written 2012-05-04

   • Added attachment scale.png.

   Sam provided this scaling plot for the new version. I think there may still be room for further improvements, but I am pleased with how it has gone so far. The coefficients caused by the creation of the tree on every processor (which will impact the initial, low-processor count runs) as well as the collective communication (which will impact the high-processor count runs) combine and give an alright result. I'll be performing my own soon.

Verification: Please write the text from the image in the box (letters only)

Is that you, Humanoid? Is this me?

• Cancel