| Message ID | pull.844.v2.git.1614123848.gitgitgadget@gmail.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Optimization batch 8: use file basenames even more | expand |
On 2/23/2021 6:43 PM, Elijah Newren via GitGitGadget wrote: > This series depends on en/diffcore-rename (a concatenation of what I was > calling ort-perf-batch-6 and ort-perf-batch-7). > > There are no changes since v1; it's just a resend a week and a half later to > bump it so it isn't lost. Thank you for re-sending. I intended to review it before but got redirected and forgot to pick it up again. > === Results === > > For the testcases mentioned in commit 557ac0350d ("merge-ort: begin > performance work; instrument with trace2_region_* calls", 2020-10-28), the > changes in just this series improves the performance as follows: > > Before Series After Series > no-renames: 12.775 s ± 0.062 s 12.596 s ± 0.061 s > mega-renames: 188.754 s ± 0.284 s 130.465 s ± 0.259 s > just-one-mega: 5.599 s ± 0.019 s 3.958 s ± 0.010 s > > > As a reminder, before any merge-ort/diffcore-rename performance work, the > performance results we started with (as noted in the same commit message) > were: > > no-renames-am: 6.940 s ± 0.485 s > no-renames: 18.912 s ± 0.174 s > mega-renames: 5964.031 s ± 10.459 s > just-one-mega: 149.583 s ± 0.751 s These are good results. I reviewed the patches and believe they do the optimizations claimed. I only found some nits for comments and whitespace things. You are very careful to create the necessary pieces and connect them from the bottom-up. However, this leads to one big "now everything is done" commit with performance improvements. It seems that there are some smaller performance improvements that could be measured if the logic was instead built from the top-down with stubs for the complicated logic. For example, the final patch links the rename logic with a call to idx_possible_rename(). But, that could just as well always return -1 and the implementation would be correct. Then, it would be good to see if the performance changes with that non-functional update. It would also help me read the series in patch order and understand the context of the methods a bit better before seeing their implementation. This is _not_ a recommendation that you rewrite the series. Just food for thought as we continue with similar enhancements in the future. Thanks, -Stolee
On Wed, Feb 24, 2021 at 9:50 AM Derrick Stolee <stolee@gmail.com> wrote: > > On 2/23/2021 6:43 PM, Elijah Newren via GitGitGadget wrote: > > This series depends on en/diffcore-rename (a concatenation of what I was > > calling ort-perf-batch-6 and ort-perf-batch-7). > > > > There are no changes since v1; it's just a resend a week and a half later to > > bump it so it isn't lost. > > Thank you for re-sending. I intended to review it before but got redirected > and forgot to pick it up again. > > > === Results === > > > > For the testcases mentioned in commit 557ac0350d ("merge-ort: begin > > performance work; instrument with trace2_region_* calls", 2020-10-28), the > > changes in just this series improves the performance as follows: > > > > Before Series After Series > > no-renames: 12.775 s ± 0.062 s 12.596 s ± 0.061 s > > mega-renames: 188.754 s ± 0.284 s 130.465 s ± 0.259 s > > just-one-mega: 5.599 s ± 0.019 s 3.958 s ± 0.010 s > > > > > > As a reminder, before any merge-ort/diffcore-rename performance work, the > > performance results we started with (as noted in the same commit message) > > were: > > > > no-renames-am: 6.940 s ± 0.485 s > > no-renames: 18.912 s ± 0.174 s > > mega-renames: 5964.031 s ± 10.459 s > > just-one-mega: 149.583 s ± 0.751 s > > These are good results. > > I reviewed the patches and believe they do the optimizations claimed. I > only found some nits for comments and whitespace things. Thanks for taking a look; I'll get those fixed up. Also, I think your performance improvement of switching from xstrfmt to a few strbuf calls, even if small, counts as more than "nits for comments and whitespace things". :-) > You are very careful to create the necessary pieces and connect them > from the bottom-up. However, this leads to one big "now everything is > done" commit with performance improvements. It seems that there are > some smaller performance improvements that could be measured if the > logic was instead built from the top-down with stubs for the complicated > logic. > > For example, the final patch links the rename logic with a call to > idx_possible_rename(). But, that could just as well always return -1 > and the implementation would be correct. Then, it would be good to see > if the performance changes with that non-functional update. It would > also help me read the series in patch order and understand the context > of the methods a bit better before seeing their implementation. > > This is _not_ a recommendation that you rewrite the series. Just food > for thought as we continue with similar enhancements in the future. I can give it a shot for future relevant patch series (some of the series this wouldn't be relevant for because they just include a collection of patches implementing separate improvements that are just batched together). A couple of the series are already structured this way, in fact, but the next series after this one has one patch that I think I could reorder to make it more like this.
This series depends on en/diffcore-rename (a concatenation of what I was calling ort-perf-batch-6 and ort-perf-batch-7). There are no changes since v1; it's just a resend a week and a half later to bump it so it isn't lost. === Optimization idea === This series uses file basenames (portions of the path after the last '/', including file extension) in a more involved fashion to guide rename detection. It's a follow-on improvement to "Optimization #3" from my Git Merge 2020 talk[1]. The basic idea behind this series is the same as the last series: people frequently move files across directories while keeping the filenames the same, thus files with the same basename are likely rename candidates. However, the previous optimization only applies when basenames are unique among remaining adds and deletes after exact rename detection, so we need to do something else to match up the remaining basenames. When there are many files with the same basename (e.g. .gitignore, Makefile, build.gradle, or maybe even setup.c, AbtractFactory.java, etc.), being able to "guess" which directory a given file likely would have moved to can provide us with a likely rename candidate if there is a file with the same basename in that directory. Since exact rename detection is done first, we can use nearby exact renames to help us guess where any given non-unique basename file may have moved; it just means doing "directory rename detection" limited to exact renames. There are definitely cases when this strategy still won't help us: (1) We only use this strategy when the directory in which the original file was found has also been removed, (2) a lack of exact renames from the given directory will prevents us from making a new directory prediction, (3) even if we predict a new directory there may be no file with the given basename in it, and (4) even if there is an unmatched add with the appropriate basename in the predicted directory, it may not meet the higher min_basename_score similarity threshold. It may be worth noting that directory rename detection at most predicts one new directory, which we use to ensure that we only compare any given file with at most one other file. That's important for compatibility with future optimizations. However, despite the caveats and limited applicability, this idea provides some nice speedups. === Results === For the testcases mentioned in commit 557ac0350d ("merge-ort: begin performance work; instrument with trace2_region_* calls", 2020-10-28), the changes in just this series improves the performance as follows: Before Series After Series no-renames: 12.775 s ± 0.062 s 12.596 s ± 0.061 s mega-renames: 188.754 s ± 0.284 s 130.465 s ± 0.259 s just-one-mega: 5.599 s ± 0.019 s 3.958 s ± 0.010 s As a reminder, before any merge-ort/diffcore-rename performance work, the performance results we started with (as noted in the same commit message) were: no-renames-am: 6.940 s ± 0.485 s no-renames: 18.912 s ± 0.174 s mega-renames: 5964.031 s ± 10.459 s just-one-mega: 149.583 s ± 0.751 s === Alternative, rejected idea === There was an alternative idea to the series presented here that I also tried: instead of using directory rename detection based on exact renames to predict where files would be renamed and then comparing to the file with the same basename in the new directory, one could instead take all files with the same basename -- both sources and destinations -- and then do a smaller M x N comparison on all those files to find renames. Any non-matches after that step could be combined with all other files for the big inexact rename detection step. There are two problems with such a strategy, though. One is that in the worst case, you approximately double the cost of rename detection (if most potential rename pairs all have the same basename but they aren't actually matches, you end up comparing twice). The second issue isn't clear until trying to combine this idea with later performance optimizations. The next optimization will provide a way to filter out several of the rename sources. If our inexact rename detection matrix is sized 1 x 4000 because we can remove all but one source file, but we have 100 files with the same basename, then a 100 x 100 comparison is actually more costly than a 1 x 4000 comparison -- and we don't need most of the renames from the 100 x 100 comparison. The advantage of the directory rename detection based idea for finding which basenames to match up, is that the cost for each file is linear (or, said another way, scales proportional to doing a diff on that file). As such, the costs for this preliminary optimization are nicely controlled and the worst case scenario is it has spent a little extra time upfront but still has to do the full inexact rename detection. [1] https://github.com/newren/presentations/blob/pdfs/merge-performance/merge-performance-slides.pdf Elijah Newren (10): Move computation of dir_rename_count from merge-ort to diffcore-rename diffcore-rename: add functions for clearing dir_rename_count diffcore-rename: move dir_rename_counts into a dir_rename_info struct diffcore-rename: extend cleanup_dir_rename_info() diffcore-rename: compute dir_rename_counts in stages diffcore-rename: add a mapping of destination names to their indices diffcore-rename: add a dir_rename_guess field to dir_rename_info diffcore-rename: add a new idx_possible_rename function diffcore-rename: limit dir_rename_counts computation to relevant dirs diffcore-rename: use directory rename guided basename comparisons Documentation/gitdiffcore.txt | 2 +- diffcore-rename.c | 439 ++++++++++++++++++++++++++++++++-- diffcore.h | 7 + merge-ort.c | 144 +---------- 4 files changed, 439 insertions(+), 153 deletions(-) base-commit: aeca14f748afc7fb5b65bca56ea2ebd970729814 Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-844%2Fnewren%2Fort-perf-batch-8-v2 Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-844/newren/ort-perf-batch-8-v2 Pull-Request: https://github.com/gitgitgadget/git/pull/844 Range-diff vs v1: 1: fec4f1d44c06 = 1: fec4f1d44c06 Move computation of dir_rename_count from merge-ort to diffcore-rename 2: 612da82f049c = 2: 612da82f049c diffcore-rename: add functions for clearing dir_rename_count 3: 93f98fc0b264 = 3: 93f98fc0b264 diffcore-rename: move dir_rename_counts into a dir_rename_info struct 4: f7bdad78219d = 4: f7bdad78219d diffcore-rename: extend cleanup_dir_rename_info() 5: 3a29cf9e526f = 5: 3a29cf9e526f diffcore-rename: compute dir_rename_counts in stages 6: dffecc064dd3 = 6: dffecc064dd3 diffcore-rename: add a mapping of destination names to their indices 7: 4983a1c2f908 = 7: 4983a1c2f908 diffcore-rename: add a dir_rename_guess field to dir_rename_info 8: cbd055ab3399 = 8: cbd055ab3399 diffcore-rename: add a new idx_possible_rename function 9: 4e095ea7c439 = 9: 4e095ea7c439 diffcore-rename: limit dir_rename_counts computation to relevant dirs 10: 1df498b3a2f0 = 10: 805c101cfd84 diffcore-rename: use directory rename guided basename comparisons