On Tue, 27 Feb 2007, Stephen Denne wrote:
> I'm sad to say that I have not created any micro-benchmark tests, and
> unfortunately the improvements are very minor, and far overshadowed by
> the variability I get from my system.
>
I've created the attached test which tests the original code (Orig), your
code (Two), and my suggestion of an int4buf (Three) and got the following
surprising results:
jurka@tony:~/pg/jdbc/projects/perf/micro$ java -classpath . Tester Orig |
sort -n
11335
11370
11468
11484
11487
jurka@tony:~/pg/jdbc/projects/perf/micro$ java -classpath . Tester Two |
sort -n
12472
12476
12489
12492
12619
jurka@tony:~/pg/jdbc/projects/perf/micro$ java -classpath . Tester Three |
sort -n
4259
4562
4564
4611
4689
This shows your code is actually slower than the original code, although I
have no idea why that could be. It shows the int4buf idea as a clear
winner. I'm a little suspicious of the whole test because of your numbers
going up. Could you take a look at this and possibly confirm the results?
I'm not sure if windows has an equivalent to /dev/null, but I wanted to
avoid any impact of disk io.
Kris Jurka