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

~bzr-pqm/bzr/bzr.dev » Changes from revision 6118.2.6

From Revision 6118.2.6 to 6105

Rev	Summary	Authors	Date
6118.2.6	Updates to ZLibEstimator. Updates to ZLibEstimator. Use the current compression ratio to estimate how much space the uncompressed tip is going to take up. This still overestimates a fair bit for gcc, but our responses are around 32kB instead of around 8kB. For bzr it seems to occasionally under-estimate, but mostly over-estimate by a bit. The only other thing I can think of, is to figure that if we're getting really high compression ratios with zlib, assume we can do even better with bz2. Note also, that gcc gets much closer to '64kB' each time because we are often only searching a small handful of tips (often even just 1), which means our granularity adding to the compressor is very tight. By comparison, bzr is often searching 1000+ tips, and we only have the option to add everything for the next step, or nothing.	John Arbash Meinel	13 years ago
6118.2.5	Disabling lsprof shows something interesting. Disabling lsprof shows something interesting. It turns out that the bulk of the slowdown when searching for more revisions was specifically the difference_update calls. As a table: time 25.2s short_search_gpm 22.2s long_search_gpm.difference 34.2s long_search_gpm.difference_update So I thought zlib.compressobj or bz2.compress was taking longer, but really it was just that our 'queried_revs' count was getting so big it was making difference_update exceedingly slow. I'm very happy that the longer searches are now even faster over the loopback than doing shorter prefetches. It makes a lot of sense, but it is nice to see.	John Arbash Meinel	13 years ago
6118.2.4	debugging. Use lsprof to determine where we are spending our... debugging. Use lsprof to determine where we are spending our time.	John Arbash Meinel	13 years ago
6118.2.3	An 'entropy' computation. An 'entropy' computation. However, pure entropy is not a good metric. Entropy for bzr revisions is about 5bits/byte, and entropy for gcc is about 4bits/byte, however compression for bzr is about 4-5:1, but for gcc it is about 20:1. This is because the compression is mostly about long substrings that are the same for gcc. So instead, we'll just try to tweak the zlib one.	John Arbash Meinel	13 years ago
6118.2.2	Avoid flushing on every loop. Gives much better estimates ... Avoid flushing on every loop. Gives much better estimates for gcc. Instead of being in the 22k range, they are now in the 50k range where we want them to be. On the downside, the computation overhead is clearly higher. I'm going to see if I can use a different algorithm for estimating compressed size.	John Arbash Meinel	13 years ago
6118.2.1	Refactor the Remote get_parent_map loop, and change how we c... Refactor the Remote get_parent_map loop, and change how we compute how much prefetch to do. Instead of estimating the compressed size by using an estimated compression ratio, use a zlib.compressobj and flush it periodically. We seem to always get better compression from the bz2 stream than the zlib.compressobj (which is expected). And it means that are estimate is rarely wildly off. At the moment, it is mostly under-estimating for bzr.dev (and for gcc), but it is doing a much closer estimate for gcc.	John Arbash Meinel	13 years ago
6118	(jelmer) Some compatibility fixes for cython.dev. (Denys Duc... (jelmer) Some compatibility fixes for cython.dev. (Denys Duchier)	Canonical.com Patch ...	13 years ago
6117	(jelmer) Print a warning rather than an error when the curre... (jelmer) Print a warning rather than an error when the current user does not have permission to read a configuration file. (Jelmer Vernooij)	Canonical.com Patch ...	13 years ago
6116	(jr) add digital signatures to log-format doc (Jonathan Ridd... (jr) add digital signatures to log-format doc (Jonathan Riddell)	Canonical.com Patch ...	13 years ago
6115	(jelmer) Move the use of hashcache to a new PreDirStateWorki... (jelmer) Move the use of hashcache to a new PreDirStateWorkingTree and case detection to InventoryWorkingTree. (Jelmer Vernooij)	Canonical.com Patch ...	13 years ago
6114	(jr) Do not run i18n initialisation twice. (Jonathan Riddell... (jr) Do not run i18n initialisation twice. (Jonathan Riddell)	Canonical.com Patch ...	13 years ago
6113	(jelmer) Properly translate IOErrors from open_file in (jelmer) Properly translate IOErrors from open_file in LocalTransport.open_write_stream. (Jelmer Vernooij)	Canonical.com Patch ...	13 years ago
6112	(vila) Fix a few missed references to bzr-2.4 (and one to py... (vila) Fix a few missed references to bzr-2.4 (and one to python-2.4) (Vincent Ladeuil)	Canonical.com Patch ...	13 years ago
6111	(jameinel) Bug #483661, (jameinel) Bug #483661, do not reuse transports for the two branches involved in a pull in order to avoid concurrent requests over a smart medium like ssh. (Martin von Gagern)	Canonical.com Patch ...	13 years ago
6110	(jelmer) Various improvements to the testsuite to allow runn... (jelmer) Various improvements to the testsuite to allow running against foreign branches. (Jelmer Vernooij)	Canonical.com Patch ...	13 years ago
6109	(vila) Better conversion for default values of config Option... (vila) Better conversion for default values of config Options. (Vincent Ladeuil)	Canonical.com Patch ...	13 years ago
6108	(mbp) top-level doc page should now point to whatsnew-2.5 (mbp) top-level doc page should now point to whatsnew-2.5	Canonical.com Patch ...	13 years ago
6107	(jameinel) Bug #827721, OS/IOError will have locale dependen... (jameinel) Bug #827721, OS/IOError will have locale dependent encoding, so %r rather than %s. (IWATA Hidetaka)	Canonical.com Patch ...	13 years ago
6106	(jameinel) Bug #835545, (jameinel) Bug #835545, lazy_import modules rather than objects to allow pydoc to work. (Jelmer Vernooij)	Canonical.com Patch ...	13 years ago
6105	(jam) Bug #388269, (jam) Bug #388269, avoid searching for missing history in the stacked branch by exposing the cache available in the stacked-on branch.	Canonical.com Patch ...	13 years ago

Older »