~bzr-pqm/bzr/bzr.dev : changes

~bzr-pqm/bzr/bzr.dev » Changes from revision 5365.5.14

Rev	Summary	Authors	Date
5365.5.14	We should be exposing StaticTuple_FromSequence in its api de... We should be exposing StaticTuple_FromSequence in its api definition	John Arbash Meinel	14 years ago
5365.5.13	Populate the offsets array. Populate the offsets array. This cuts down the number of bisections dramatically, basically by pre-caching the first step. On real-world data it drops the steps from 587 to 156. Or from 4.9/key to 1.3/key. This drops the time to lookup from 23.7us to 20.3us. Note that (k in dict) is 12.2us. I do wish we were just a bit closer to that. However, with _LeafNode inherited from dict, I get 26us, so maybe there is something in the interpreter that does a PyDict_CheckExact call, and there isn't much we can do about it.	John Arbash Meinel	14 years ago
5365.5.12	Make _LeafNode inherit from dict (is-a rather than have-a) Make _LeafNode inherit from dict (is-a rather than have-a) It avoids having another object layer, and indirection via attribute and forwarding functions. TIMEIT shows: 14.6us raw dict 26.6us subclass 57.9us wrapper class	John Arbash Meinel	14 years ago
5365.5.11	get into the nitty gritty for the _key_to_sha1 function. get into the nitty gritty for the _key_to_sha1 function. It doesn't seem to help to op-out of the actual unhexlify call, so at least that doesn't seem to be the performance overhead. We get down to around 19.9us for _key_to_sha1 across 120 keys, which is 0.166us per key.	John Arbash Meinel	14 years ago
5365.5.10	Do some timing with different objects. Do some timing with different objects. It turns out that pure dict is 13.7us lookup time, dict-as-attr is 20.7us, PyObject __getitem__ to dict-as-attr is 54us.	John Arbash Meinel	14 years ago
5365.5.9	Add some tests that key lookup works. Add some tests that key lookup works. Timing shows that looking up all 120 keys takes 205us, using bisect takes 184us. So better, but not great. I'm a bit surprised it isn't faster, but perhaps the comparison is a lot less of total time than the conversion to python objects.	John Arbash Meinel	14 years ago
5365.5.8	Add tests that we handle sizes close to and >2GB properly. Add tests that we handle sizes close to and >2GB properly. We can't use %llu because that wasn't added until python 2.7, but we can use str(PyLong).	John Arbash Meinel	14 years ago
5365.5.7	Document some perf decisions. Document some perf decisions.	John Arbash Meinel	14 years ago
5365.5.6	For a single page test, we parse() still takes 47us For a single page test, we parse() still takes 47us parse().all_keys() takes 66.6us if we don't intern, and 86.9us if we do. (10us is hash() time) We'll need to think about whether interning is worthwhile. For some cases we are optimizing for, it will bloat. However, if we only do keys that we actually return, then likely we are ok, and ahead to intern. Because those keys will then get used to read chk maps, or have been read from them. So I'll leave intern() in for now.	John Arbash Meinel	14 years ago
5365.5.5	Found some more bugs. I wasn't incrementing one of the point... Found some more bugs. I wasn't incrementing one of the pointers, so I was parsing all the records, but not saving them correctly.	John Arbash Meinel	14 years ago
5365.5.4	Some direct tests of the hex and unhex functions. Some direct tests of the hex and unhex functions. As they say, untested code is broken code. And my code was broken. (I didn't actually init the array, I had bugs in my init strings, etc.)	John Arbash Meinel	14 years ago
5365.5.3	The results were just too promising. faster and better mem... The results were just too promising. faster and better memory.	John Arbash Meinel	14 years ago
5365.5.2	Parse times are dramatically improved. Parse times are dramatically improved. using _LeafNode directly, the total parse time is actually 369us. Putting all the items into a 'dict' ends up costing 250us by itself, and a minor overhead to put that into the _LeafNode object itself. Switching it around to call parsed(buffer...).all_items() to actually construct the items again, and we see that it costs another 120us for the _LeafNode implementation (which has to sort), while the Pyrex implementation goes from 50us to parse, to just 120us to create all the objects and return them. It doesn't have to sort, because they are stored in sorted order. all_keys() is even better at 94us vs 454us. I'm rather impressed that PyString_FromFormat is actually pretty reasonable. Certainly inefficient overall (since we'll eventually parse back into ints again), but does a good job.	John Arbash Meinel	14 years ago
5365.5.1	Implement a custom parser for chk btree leaves. Implement a custom parser for chk btree leaves. The basic parsing is actually quicker (47us to build this new structure vs 110us) Most likely because we have 1 malloc rather than N per leaf. Memory size is also much smaller. We'll lose a little bit if we convert back and forth from keys, etc.	John Arbash Meinel	14 years ago
5365	(Andrew Bennetts) Merge 2.2 branch back into trunk (Andrew Bennetts) Merge 2.2 branch back into trunk	Canonical.com Patch ...	14 years ago
5364	[r=jam] Add VersionedFiles.get_annotator() stub. (Jelmer) [r=jam] Add VersionedFiles.get_annotator() stub. (Jelmer)	Canonical.com Patch ...	14 years ago
5363	(mbp) delete deprecated RootEntry class and InventoryEntry m... (mbp) delete deprecated RootEntry class and InventoryEntry methods (Martin Pool)	Canonical.com Patch ...	14 years ago
5362	(spiv) Expand lp:~/ to lp:~username/ if a user has already l... (spiv) Expand lp:~/ to lp:~username/ if a user has already logged in. (John A Meinel)	Canonical.com Patch ...	14 years ago
5361	(garyvdm, gz) Hack py2exe so that we can have a library.zip ... (garyvdm, gz) Hack py2exe so that we can have a library.zip built with optimize=2, but every thing else has optimize=1.	Canonical.com Patch ...	14 years ago
5360	(gz) Correct selftest bzr script location lookup logic when ... (gz) Correct selftest bzr script location lookup logic when bzrlib is in the cwd (Martin [gz])	Canonical.com Patch ...	14 years ago

Older »