Hi folks,
I finally have round two of my perf work ready. Please refer to
http://defect.opensolaris.org/bz/show_bug.cgi?id=2344#c1 for the
full perf results, but the punchline is:
OLD CODE NEW CODE DIFF
----------------------------------------------------------------------
pkg image-update -n 130.6 sec 103.4 sec 1.2x
pkg list 88.8 sec 15.6 sec 5.6x
pkg list -a 106.8 sec 32.5 sec 3.2x
pkg install -n sunstudio 40.4 sec 32.5 sec 1.2x
Please look this over carefully. I would like at least two reviewers.
http://cr.opensolaris.org/~dp/fmri-perf
Here are the major changes:
DotSequence:
- Subclass of list rather than a wrapper around one. This cuts the
storage needed for an FMRI by roughly 50%, from 1150 bytes to about
550. 17% slower to create when done this way (but other operations
are much faster).
- Simplified greatly since lists just "do the right thing" for
comparison, and seem to be a lot faster, see also bug 2363.
Version:
- Dropped support for pre-ISO timestamps
- Cleanup of gt and lt
- Timestamps now stored as strings.
FMRI:
- 25x speedup in hash implementation
- Cleanup gt and lt
As part of this, I have developed fmribench.py, which I will also
be checking in. Also keep in mind that while the perf gains are
nice, as the number of known package versions continues to grow,
performance will continue to suffer, until we get the algorithmic
improvements which Danek has been working on. My hope is that this
work will help keep the dam from breaking until that point.
I would also like to share what was for me the big lesson here:
think hard when you implement:
__hash__
__eq__, __lt__, etc.
__str__
Thanks,
-dp
--
Daniel Price - Solaris Kernel Engineering - [EMAIL PROTECTED] - blogs.sun.com/dp
_______________________________________________
pkg-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/pkg-discuss
