As a test for overall performance of BigDecimal, I took a short piece of code for calculating the number 'e' by serial expansion from a posting by Michael Quinlan. I am not convinced this code is correct (or at least carefull enough) in matters of convergence, but I checked the 3 classes to return exactly the same result up to 2000 digits, so for performance measurement this algorithm will do.

 /**Calculates the number 'e' with the requested precision  
  * @param iPrecision the requested precision
  * @param bdOne BigDecimal '1', used for differentating this method's signatures 
  *              when comparing different BigDecimal implementations
  */
 static String numberE(int iPrecision, be.arci.math.BigDecimal bdOne)
 {
  be.arci.math.BigDecimal curfact = bdOne;
  be.arci.math.BigDecimal factmul = bdOne;
  be.arci.math.BigDecimal curval = be.arci.math.BigDecimal.ZERO;
  String sE = "";
  
  for (int iIteration = 0; true; iIteration++)
  {
   //System.out.print("Iteration " + iIteration + "\r");
   // divide 1 by the current factorial
   be.arci.math.BigDecimal bdTerm = bdOne.divide(curfact, 
                                                 iPrecision + 1, 
                                                 be.arci.math.BigDecimal.ROUND_HALF_EVEN);
   // add the result to the accumulated value
   curval = curval.add(bdTerm); 
   // check convergence of the current value
   String s = curval.toString().substring(0, iPrecision + 2);
   if (s.equals(sE)) 
   {
    //System.out.println("");
    //System.out.println(s);
    break;
   }
   sE = s;
   // move to the next factorial value
   curfact = curfact.multiply(factmul);
   factmul = factmul.add(bdOne);
  }
  return sE;
 }