Changes between Version 9 and Version 10 of ScramblerPaper

Timestamp:

05/11/11 14:48:53 (14 years ago)

Author:

Jonathan

Comment:

—

ScramblerPaper

@@ -41 +41 @@
 
 * Load Balancing
- * As was mentioned before, threading level removes the need for balancing each level and instead allows for global load balancing.  During each root step (level 0) there will be 1 set of level 1 grids, R sets of level 2 grids, R^2^ sets of level 3 grids, and R^n-1^ sets of level n grids that will need to be distributed where R is the refinement level.  The basic idea is that before each new distribution of level n grids, each processor evaluates how much work remains to be finished on levels coarser then n (0:n-1) per level n-1 step as well as the new workload associated with its new children over one n-1 level advance (or R level n advances).  These two numbers are then shared among all processors so that each processor can determine its share of the next level grids so that if the amr patch was static, each processor would finish updating every coarser grid at the end of the last level n step.  
+ * As was mentioned before, threading level removes the need for balancing each level and instead allows for global load balancing.  When distributing children of level $l$ nodes each processor first calculates the average remaining workload that needs to be accomplished per average level $l$ step.  If there are $n_l$ steps remaining before level $m$ needs to finish its updates then the workload on level $m$ is divided by $n_l$ 
 
-Basically since the coarser grid updates can be done at anytime we would like to  
+During each root step (level 0) there will be 1 set of level 1 grids, R sets of level 2 grids, R^2^ sets of level 3 grids, and R^n-1^ sets of level n grids that will need to be distributed where R is the refinement level.  The basic idea is that before each new distribution of level n grids, each processor evaluates how much work remains to be finished on levels coarser then n (0:n-1) per level n-1 step as well as the new workload associated with its new children over one n-1 level advance (or R level n advances).  These two numbers are then shared among all processors so that each processor can determine its share of the next level grids so that if the amr patch was static, each processor would finish updating every coarser grid at the end of the last level n step.  
 
 {{{