Re: [lmi] Ledger::write() -- adding DOCTYPE support to xmlwrapp?

[Greg Chicares]

On 2008-12-13 15:36Z, Vaclav Slavik wrote:
> 
> On Fri, 2008-12-12 at 02:55 +0000, Greg Chicares wrote:
> 
>> The only thing I know for sure is
>> that I measured a 25% performance penalty as noted in that comment.
> 
> Do you remember how to run this particular test?

Yes, I think so. Run the 'input' unit test (with lmi's makefiles:
  make unit_tests unit_test_targets=input_test.exe
or the autotools equivalent) before and after applying this
one-character patch:

Index: input_xml_io.cpp
===================================================================
RCS file: /sources/lmi/lmi/input_xml_io.cpp,v
retrieving revision 1.10
diff -U 3 -r1.10 input_xml_io.cpp
--- input_xml_io.cpp    17 Oct 2008 22:55:34 -0000      1.10
+++ input_xml_io.cpp    13 Dec 2008 16:42:50 -0000
@@ -135,7 +135,7 @@
 // twenty-five percent slower. What would be really desirable is an
 // (efficient) element-iterator class.
 //
-#if 0
+#if 1
     xml_lmi::ElementContainer const elements(xml_lmi::child_elements(x));
     typedef xml_lmi::ElementContainer::const_iterator eci;
     for(eci i = elements.begin(); i != elements.end(); ++i)

Here are my results (I discarded the first iteration each way,
in order to stabilize the timings):

before:
  Overhead: 3.131e-005 s =      31307 ns, mean of 320 iterations
  Vector  : 2.280e-003 s =    2279796 ns, mean of 100 iterations
  Read    : 8.161e-003 s =    8160669 ns, mean of 100 iterations
  Write   : 2.923e-003 s =    2923109 ns, mean of 100 iterations
  'cns' io: 5.782e-002 s =   57818859 ns, mean of  18 iterations
  'ill' io: 1.504e-002 s =   15037891 ns, mean of  67 iterations

after:
  Overhead: 3.215e-005 s =      32152 ns, mean of 312 iterations
  Vector  : 2.232e-003 s =    2232184 ns, mean of 100 iterations
  Read    : 1.020e-002 s =   10199924 ns, mean of  99 iterations
  Write   : 2.869e-003 s =    2869494 ns, mean of 100 iterations
  'cns' io: 6.154e-002 s =   61541580 ns, mean of  17 iterations
  'ill' io: 1.738e-002 s =   17377635 ns, mean of  58 iterations

ratio, after/before:
  Overhead: 103%
  Vector  : 98%
  Read    : 125%  <-- There's the 25% difference.
  Write   : 98%
  'cns' io: 106%
  'ill' io: 116%