Performance metrics and bug report

Jason has probably already run much better experiments, but I ran some
quick unscientific experiments with ifile and figured I'd share the
results.  Executive summary:

- running ifile -c -q (query only) on a large number of messages
  (1040) has a speedup factor of about 40 (that's four-zero) per-word
  over running it on a single message; even going to 32 messages, you
  get a speedup of about 10.

- running ifile -c -Q (query and update) or ifile -i (insert)
  on 32 messages you get a  speedup of about 15 over running it
  on a single message.

  BUG REPORT: Going to 1040 messages, ifile dumps core: ifile_realloc
  returns a 0 which is then used without being checked in
  ifile_lexer_simple_open_text_fp.  This actually explains why, in
  spite of training ifile on my mail spool, some stuff was never
  getting filed right; the initial training dumped core on the largest
  directories of mail before writing back the db ;-)


In my environment, ifile is run at least twice for every message, and
tends to take 1.5-2 seconds per message.  So daemonizing ifile might
get me down to .1 seconds/message or so; that would be a huge
improvement, especially when refiling mail.

Jason, the bug above aside, how clean is ifile with respect to memory
allocation/cleanup/leaks?

Jeremy



------------------------------------------------------------
Experimental setup: All were run with ifile 1.0.7 on on a 600Mhz
Alpha, with files stored remotely in NFS.  Performance on a faster
processor and a local filesystem would probably be a lot better, but
this is the harsh environment in which I actually read mail, so...

All timings are from 'time', and they're generally my eyeball-average
of between one and three runs.  Did I mention this was unscientific?


------------------------------------------------------------
------------------------------------------------------------
ifile -c -q  (query only)
------------------------------------------------------------
Experiment:
- ifile -c -q ARG  > /dev/null
  run on 1 message with (according to wc) 644 words

Realtime:
- Total: .541s
- Average per-word:  8.40e-4s

------------------------------------------------------------
Experiment:
- ifile -c -q ARGS  > /dev/null
  run on 32 messages with (according to wc) a total of 10712 words,
  (324.6  words-per-message on average)

Realtime:
- Total: .756s
- Average per-word: 7.06e-05s
- Average per-message:  2.29e-02s

------------------------------------------------------------
Experiment:
- ifile -c -q ARGS  > /dev/null
  run on 1070 messages with (according to wc) a total of 447770 words,
  (418.5  words-per-message on average)

Realtime:
- Total: 8.817s
- Average per-word:  1.97e-5
- Average per-message:  8.27e-3s

------------------------------------------------------------
------------------------------------------------------------
ifile -c -Q (query and update)
(ifile -i takes about the same running times)
------------------------------------------------------------
Experiment:
- ifile -c -Q ARG  > /dev/null
  run on 1 message with (according to wc) 644 words

Realtime:
- Total: 1.17s
- Average per-word:  1.82e-3

------------------------------------------------------------
Experiment:
- ifile -c -Q ARGS  > /dev/null
  run on 32 messages with (according to wc) a total of 10712 words,
  (324.6  words-per-message on average)

Realtime:
- Total: 1.25s
- Average per-word: 1.17e-4
- Average per-message:  3.90e-2

------------------------------------------------------------
Experiment:
- ifile -c -Q ARGS  > /dev/null
  run on 1070 messages with (according to wc) a total of 447770 words,
  (418.5  words-per-message on average)

Realtime:
- Total: highly variable -- 37.77 to 51.83s.  Average: 44.8s
- Average per-word:  1.00e-4
- Average per-message:  4.19e-2

NOTE: SEGMENTATION FAULT ENDS THIS EXPERIMENT; THESE RESULTS AREN'T
RELIABLE.

------------------------------------------------------------

End of data.