Subject: Algorithm efficiency and computer strength Posted: 8/22/2012 Viewed: 27131 times
I have a model that takes 15 minutes to run.
Therefore, I would like to upgrade/buy a new computer mainly because of that model.
The question is, in which component should I invest to improve performence most: Faster clock speed, multiple core, CPU cache, more RAM, faster RAM, faster hard-drive, 64bit OS (windows 7), something else?
Another question is how to build large model with hundreds of consumers.
How much faster will the model run if I group the consumers and put all the data in branches underneath the groups?
For the sake of argument let's say I have 150 consumers that I group into 40 groups.
Gilad
Mr. Jack Sieber
Subject: Re: Algorithm efficiency and computer strength Posted: 8/22/2012 Viewed: 27115 times
In terms of computer hardware, the two most important factors for speeding up WEAP calculations are CPU speed and hard disk speed. (WEAP can only use a single CPU core, but having at least two CPU cores will allow WEAP to not have to share a core with other processes.)
Replacing a hard disk with a SSD (Solid State Drive) will yield substantial improvement in calculation time for large WEAP models. However, the benefit will be even larger if you create a RAMDisk on your computer and tell WEAP to use it for its working directory. (See http://www.weap21.org/index.asp?action=9&read=1352&fID=30 for information on how to do this, and suggestions for RAMDisk software.) However, in order to use a RAMDisk, you must have enough RAM on your computer to create a RAMDisk large enough to hold your WEAP dataset, including results files. 4 GB total RAM may or may not be enough; 8 GB should be enough.
In terms of WEAP models, calculation time will be roughly proportional to the number of elements on the schematic. For example having 150 demand sites will be slower than 40 demand sites. Therefore, by grouping several demands into a fewer demand sites will improve the time. Also, it will be proportional to the number of scenarios to calculate and the number of years and timesteps. A weekly simulation will take much longer than a monthly simulation. You should balance how important it is to have the extra detail (more demand sites or smaller timesteps) vs. the extra time and disk space used. No model can perfectly capture reality--the trick is in finding a good balance between having realistic detail and what is realistic to model.
Jack
Topic "Algorithm efficiency and computer strength"