171 |
|
|
172 |
|
#ifdef IS_MPI |
173 |
|
|
174 |
+ |
/********************************************************************* |
175 |
+ |
* Documentation? You want DOCUMENTATION? |
176 |
+ |
* |
177 |
+ |
* Why all the potatoes below? |
178 |
+ |
* |
179 |
+ |
* To make a long story short, the original version of DumpWriter |
180 |
+ |
* worked in the most inefficient way possible. Node 0 would |
181 |
+ |
* poke each of the node for an individual atom's formatted data |
182 |
+ |
* as node 0 worked its way down the global index. This was particularly |
183 |
+ |
* inefficient since the method blocked all processors at every atom |
184 |
+ |
* (and did it twice!). |
185 |
+ |
* |
186 |
+ |
* An intermediate version of DumpWriter could be described from Node |
187 |
+ |
* zero's perspective as follows: |
188 |
+ |
* |
189 |
+ |
* 1) Have 100 of your friends stand in a circle. |
190 |
+ |
* 2) When you say go, have all of them start tossing potatoes at |
191 |
+ |
* you (one at a time). |
192 |
+ |
* 3) Catch the potatoes. |
193 |
+ |
* |
194 |
+ |
* It was an improvement, but MPI has buffers and caches that could |
195 |
+ |
* best be described in this analogy as "potato nets", so there's no |
196 |
+ |
* need to block the processors atom-by-atom. |
197 |
+ |
* |
198 |
+ |
* This new and improved DumpWriter works in an even more efficient |
199 |
+ |
* way: |
200 |
+ |
* |
201 |
+ |
* 1) Have 100 of your friend stand in a circle. |
202 |
+ |
* 2) When you say go, have them start tossing 5-pound bags of |
203 |
+ |
* potatoes at you. |
204 |
+ |
* 3) Once you've caught a friend's bag of potatoes, |
205 |
+ |
* toss them a spud to let them know they can toss another bag. |
206 |
+ |
* |
207 |
+ |
* How's THAT for documentation? |
208 |
+ |
* |
209 |
+ |
*********************************************************************/ |
210 |
+ |
|
211 |
|
int *potatoes; |
212 |
|
int myPotato; |
213 |
|
|