| Submitter | Jun Wu |
|---|---|
| Date | Sept. 2, 2016, 3:10 p.m. |
| Message ID | <c9a9040c95add098eb8e.1472829030@x1c> |
| Download | mbox | patch |
| Permalink | /patch/16533/ |
| State | Changes Requested |
| Headers | show |
Comments
On Fri, 2 Sep 2016 16:10:30 +0100, Jun Wu wrote: > # HG changeset patch > # User Jun Wu <quark@fb.com> > # Date 1472826059 -3600 > # Fri Sep 02 15:20:59 2016 +0100 > # Node ID c9a9040c95add098eb8e280fa60edaa766154141 > # Parent d130a38ef41f3c9e2d2f26df3535d89a93f87301 > # Available At https://bitbucket.org/quark-zju/hg-draft > # hg pull https://bitbucket.org/quark-zju/hg-draft -r c9a9040c95ad > annotate: pre-calculate the "needed" dictionary > > The "needed" dict is used as a reference counter to free items in the giant > "hist" dict. However, currently it is not very accurate and can lead to > dropping "hist" items unnecessarily, for example, with the following DAG, > > -3- > / \ > --1--2--4-- > > The current algorithm will visit and calculate rev 1 twice, undesired. Also that would generate wrong result probably because hist[f] was deleted too early and pcache[f] was emptied. Compare "hg annotate mercurial/commands.py" for example. Can you add a test case for that? This patch appears to fix old bug, so I think it's good for stable. > + # 1st DFS pre-calculates pcache and needed > visit = [base] > - hist = {} > pcache = {} > needed = {base: 1} > while visit: > f = visit[-1] > - pcached = f in pcache > - if not pcached: > - pcache[f] = parents(f) > + visit.pop() Nit: could be f = visit.pop() > + pl = parents(f) > + pcache[f] = pl > + for p in pl: > + if p not in needed: > + needed[p] = 1 > + else: > + needed[p] += 1 > + if p not in pcache: > + visit.append(p) There can be more than one route to 'p' from 'visit[]', so 'p not in pcache' isn't enough to avoid unneeded calculation of 'parents(f)', and needed[p] can be over-counted.
Good catch about both issues! I will prepare a V2. Excerpts from Yuya Nishihara's message of 2016-09-05 00:21:45 +0900: > On Fri, 2 Sep 2016 16:10:30 +0100, Jun Wu wrote: > > # HG changeset patch > > # User Jun Wu <quark@fb.com> > > # Date 1472826059 -3600 > > # Fri Sep 02 15:20:59 2016 +0100 > > # Node ID c9a9040c95add098eb8e280fa60edaa766154141 > > # Parent d130a38ef41f3c9e2d2f26df3535d89a93f87301 > > # Available At https://bitbucket.org/quark-zju/hg-draft > > # hg pull https://bitbucket.org/quark-zju/hg-draft -r c9a9040c95ad > > annotate: pre-calculate the "needed" dictionary > > > > The "needed" dict is used as a reference counter to free items in the giant > > "hist" dict. However, currently it is not very accurate and can lead to > > dropping "hist" items unnecessarily, for example, with the following DAG, > > > > -3- > > / \ > > --1--2--4-- > > > > The current algorithm will visit and calculate rev 1 twice, undesired. > > Also that would generate wrong result probably because hist[f] was deleted too > early and pcache[f] was emptied. Compare "hg annotate mercurial/commands.py" > for example. > > Can you add a test case for that? This patch appears to fix old bug, so I think > it's good for stable. > > > + # 1st DFS pre-calculates pcache and needed > > visit = [base] > > - hist = {} > > pcache = {} > > needed = {base: 1} > > while visit: > > f = visit[-1] > > - pcached = f in pcache > > - if not pcached: > > - pcache[f] = parents(f) > > + visit.pop() > > Nit: could be f = visit.pop() > > > + pl = parents(f) > > + pcache[f] = pl > > + for p in pl: > > + if p not in needed: > > + needed[p] = 1 > > + else: > > + needed[p] += 1 > > + if p not in pcache: > > + visit.append(p) > > There can be more than one route to 'p' from 'visit[]', so 'p not in pcache' > isn't enough to avoid unneeded calculation of 'parents(f)', and needed[p] can > be over-counted.
Patch
diff --git a/mercurial/context.py b/mercurial/context.py --- a/mercurial/context.py +++ b/mercurial/context.py @@ -991,13 +991,26 @@ class basefilectx(object): # depth-first search. + # 1st DFS pre-calculates pcache and needed visit = [base] - hist = {} pcache = {} needed = {base: 1} while visit: f = visit[-1] - pcached = f in pcache - if not pcached: - pcache[f] = parents(f) + visit.pop() + pl = parents(f) + pcache[f] = pl + for p in pl: + if p not in needed: + needed[p] = 1 + else: + needed[p] += 1 + if p not in pcache: + visit.append(p) + + # 2nd DFS does the actual annotate + visit[:] = [base] + hist = {} + while visit: + f = visit[-1] ready = True @@ -1007,6 +1020,4 @@ class basefilectx(object): ready = False visit.append(p) - if not pcached: - needed[p] = needed.get(p, 0) + 1 if ready: visit.pop()
# HG changeset patch # User Jun Wu <quark@fb.com> # Date 1472826059 -3600 # Fri Sep 02 15:20:59 2016 +0100 # Node ID c9a9040c95add098eb8e280fa60edaa766154141 # Parent d130a38ef41f3c9e2d2f26df3535d89a93f87301 # Available At https://bitbucket.org/quark-zju/hg-draft # hg pull https://bitbucket.org/quark-zju/hg-draft -r c9a9040c95ad annotate: pre-calculate the "needed" dictionary The "needed" dict is used as a reference counter to free items in the giant "hist" dict. However, currently it is not very accurate and can lead to dropping "hist" items unnecessarily, for example, with the following DAG, -3- / \ --1--2--4-- The current algorithm will visit and calculate rev 1 twice, undesired. However, simply removing "needed" is not okay, because it will consume 10x memory: # without any change % HGRCPATH= lrun ./hg annotate mercurial/commands.py -r d130a38 3>&2 [1] MEMORY 49074176 CPUTIME 9.213 REALTIME 9.270 # with "needed" removed MEMORY 637673472 CPUTIME 8.164 REALTIME 8.249 This patch moves "needed" (and "pcache") calculation to a separate DFS to address the issue. It improves perf (reuses hist) while maintaining low memory usage: # with this patch applied MEMORY 48164864 CPUTIME 7.871 REALTIME 7.956 [1]: lrun is a lightweight sandbox built on Linux cgroup and namespace. It's used to measure CPU and memory usage here. Source code is available at github.com/quark-zju/lrun.