[XML-SIG] New data on speed of string appending

Jeremy Hylton jeremy@beopen.com
Thu, 24 Aug 2000 12:36:14 -0400 (EDT)

>>>>> "TP" == tpassin  <tpassin@home.com> writes:

  TP> Remember back in June (June 22,2000), Bjorn Pettersen showed how
  TP> using StringIO instead of just appending to a string was
  TP> **much** faster?  Then others showed that using
  TP> list.append/join(list) was also very fast.


  TP> The results are dramatic.  Method 1) is as good as or better
  TP> than anything until the string length exceeds about 1000 bytes.
  TP> Then Method 1 starts slowing down.  Above about 4000 bytes, it's
  TP> really getting ssslllooowww.  Here is a table of the results on
  TP> my system - 450 MHz PIII running Win98, Python 1.5.2.

This is an empirical confirmation of something that analysis shows
clearly.  The repeated string concatenation method (buf = buf + s)
creates a new string object each time; this costs a malloc and two
memcpys.  As the string increases, the begin of the string is copied
repeatedly.  The other methods defer the creation of a new string
object until it is needed; it defers the mallocs and memcpy until the
end and does them once.