Measurements on Checkpointing

From: sichiwai (Christian.M.Iwainsky_at_informatik.stud.uni-erlangen.de)
Date: Thu Mar 16 2006 - 17:36:52 PST

  • Next message: RETScreen International: "Clean Energy Decision Support Centre / Centre d'aide � la d�cision sur les �nergies propres"
    Hello,
     I did some performance measurements of the time that blcr takes to
    complete a checkpoint and compared it to the time a programm needs to
    write the same amount of data from memory to the harddrive. The same
    behaviour can reproduced and can be observed on multiple runs.
    
    I measured on a Intel(R) Xeon(TM) CPU 2.66GHz with 2Gig of RAM.
    
    When I plotted the performance of BLCR, I discovered "erratic" behaviour
    from 700 to 900 megabytes of checkpoint data.
    As you stated in your papers, blcr performance drops when approximately
    half of the memory of the machine is used, due to the VMADump.
    Maybe you have some insight what might cause this behaviour?
    
    I attached the graph as an PDF. The X-Axis is the amount of data to
    checkpoint/write and the Y-Axis is the time the program took to
    checkpoint/write the data to local storage. (Sorry for german labels)
    
    Greetings
     Christian
    
    


  • Next message: RETScreen International: "Clean Energy Decision Support Centre / Centre d'aide � la d�cision sur les �nergies propres"