On Wednesday, May 1, 2024 6:01 PM, Jeff King wrote: >On Tue, Apr 30, 2024 at 10:26:32PM -0700, Dhruva Krishnamurthy wrote: > >> Cloning by specifying depth exhibits performance regression in >> pack-objects (~20x). The repository I am cloning is on NFS (mounted >> with NFSv3 & positive lookup cache enabled). >> >> Ran under 'perf' command to capture profiling information to see if >> something really stands out. There is a significant overhead in calls >> to file open/open64, fstat64 & mmap/munmap in git 2.44 compared to git >> 2.42. Not sure if there is an increase in the number of calls or >> something more is done. >> >> Could someone please guide me on how to troubleshoot this better? > >Oof. I was able to reproduce this regression, and the impact can be pretty severe. >You can reproduce it without NFS and without shallow clones. Get a bare clone of >something big like the kernel: > > git clone --bare https://github.com/torvalds/linux > cd linux.git > >and then do something similar to the server side of a clone: > > time git pack-objects --all --stdout </dev/null >/dev/null > >Most of the time will be spent in the "Enumerating objects" phase, walking the >graph (since the repo is fully packed, delta searching and writing are fairly quick). > >With v2.42.0, I get: > > real 1m12.409s > user 1m13.021s > sys 0m6.091s > >With v2.44.0, I get: > > real 4m12.075s > user 4m12.763s > sys 0m5.546s > >Bisecting show the culprit is 2386535511 (attr: read attributes from HEAD when >bare repo, 2023-10-13), which is in v2.43.0. Before that, a bare repository would >only look for attributes in the info/attributes file. But after, we look at the HEAD >tree-ish, too. And pack-objects will check the "delta" attribute for every path of >every object we are packing. And remember that in-tree lookups for foo/bar/baz >require looking not just for .gitattributes, but also foo/.gitattributes, >foo/bar/.gitattributes, and foo/bar/baz/.gitattributes. > >In a non-bare repo, this isn't _too_ bad. We'll try an open() for each path, but the >common case is that we don't have such a file, so the total cost is the one syscall per >object. But when we're pulling attributes from HEAD, each lookup requires walking >the whole chain of trees (so for the final one, tree foo points to tree bar points to >tree baz, where we see there is no .gitattributes entry). And so all of that extra time >goes to reading in trees over and over. > >You can repeat the same test with git.git. It gets slower, but it's barely perceptible. >This is because we have a relatively shallow tree, so most lookups are hitting root- >level .gitattributes or maybe one or two levels deep. The kernel has a much deeper >tree, so those lookups are more expensive (and of course there are a lot more of >them). > >Getting back to shallow clones and NFS: > > - the effect is more pronounced on NFS because the cost to access > objects is higher. With git.git I was even able to see some > slowdown, probably just from the cost/latency of opening up all of > those objects. > > - the problem doesn't show up if the repo has reachability bitmaps. > This is because the bitmap result doesn't have the pathnames of each > object (we do have the "name hash", but it's not enough for us to do > an attr lookup), and so objects we get from a bitmap do not > respect the delta attribute at all. > > But when doing a shallow clone, we have to disable bitmaps and do a > regular traversal. So even if you have bitmaps, you still run into > the problem. > > The example above should not have bitmaps (we do build them by > default when repacking bare repos these days, but I don't think > we'll do so right after cloning). If you have a local repo that > already has bitmaps, you should be able to see the difference by > using "git -c pack.usebitmaps=false pack-objects". > > So even if you are a server which generally enables bitmaps, you can > still get bit by this for shallow clones, but also for other > non-bitmap invocations, like say "git repack -ad". There I see the > same 3-minute slowdown in the enumeration phase. > >So what to do? It seems like some kind of caching would help here. We're looking >up the same paths over and over, for two reasons: > > 1. We'll have many objects with the same paths, one for each time the > path was modified through history. > > 2. Adjacent objects share the higher-level lookups. Both "dir/a" and > "dir/b" will need to look up "dir/.gitattributes" (and all the way > up to ".gitattributes"). > >So even something simple and stupid like this: > >diff --git a/attr.c b/attr.c >index 679e42258c..b32af9a78b 100644 >--- a/attr.c >+++ b/attr.c >@@ -24,6 +24,7 @@ > #include "thread-utils.h" > #include "tree-walk.h" > #include "object-name.h" >+#include "strmap.h" > > const char *git_attr_tree; > >@@ -795,25 +796,31 @@ static struct attr_stack *read_attr_from_blob(struct >index_state *istate, > const struct object_id *tree_oid, > const char *path, unsigned flags) { >- struct object_id oid; >- unsigned long sz; >- enum object_type type; >+ static struct strmap cache = STRMAP_INIT; >+ void *CACHE_MISSING = (void *)1; > void *buf; >- unsigned short mode; > > if (!tree_oid) > return NULL; > >- if (get_tree_entry(istate->repo, tree_oid, path, &oid, &mode)) >- return NULL; >- >- buf = repo_read_object_file(istate->repo, &oid, &type, &sz); >- if (!buf || type != OBJ_BLOB) { >- free(buf); >- return NULL; >+ buf = strmap_get(&cache, path); >+ if (!buf) { >+ struct object_id oid; >+ unsigned short mode; >+ if (!get_tree_entry(istate->repo, tree_oid, path, &oid, &mode)) { >+ unsigned long sz; >+ enum object_type type; >+ buf = repo_read_object_file(istate->repo, &oid, &type, >&sz); >+ if (!buf || type != OBJ_BLOB) { >+ free(buf); >+ buf = NULL; >+ } >+ } >+ strmap_put(&cache, path, buf ? buf : CACHE_MISSING); > } > >- return read_attr_from_buf(buf, path, flags); >+ return buf == CACHE_MISSING ? NULL : >+ read_attr_from_buf(buf, path, flags); > } > > static struct attr_stack *read_attr_from_index(struct index_state *istate, > >restores the v2.42 performance. But there are probably better options: > > - this is caching whole .gitattributes buffers. In pack-objects we > only care about a single bit for try_delta. For linux.git it doesn't > really matter, as 99% of our entries are just CACHE_MISSING and the > real value is avoiding the negative lookups. But the same problem > exists to a lesser degree for "git log -p" in a bare repo. So I > think it makes sense to try to solve it in the attr layer. > > - the string keys have a lot of duplication. You'll have > "foo/.gitattributes", "foo/bar/.gitattributes", and so on. A trie > structure split by path component would let you store each component > just once. And perhaps have even faster lookups. I think this is > roughly the same issue faced by the kernel VFS for doing path > lookups, so something dentry/dcache-like would help. > > I don't know how much it matters in practice, though. The sum of all > of the paths in HEAD for linux.git is ~3.5MB, which is a rounding > error on the needs of the rest of the packing process. > > - Something dcache-like could also be pushed down lower, to the > get_tree_entry() API (imagine it quietly caching uncompressed trees > behind the scenes and then using them to traverse). That depends on > high locality of requests, though. I don't know if we have that > here, because we do our lookups in traversal order. So you'd look at > entries in HEAD^{tree}, then HEAD~1^{tree}, then HEAD~2^{tree}, and > so on. > > - Speaking of locality, the attr code tries to make use of request > locality in its stack (so if we ask for attributes for "foo/bar", > then "foo/baz", we should be able to keep the parsed data for > "foo/.gitattributes" available between them). But our pattern here > violates that. Again, not that big an issue if you don't have that > many .gitattributes files in the first place, so it might not be > worth worrying about. But if we did want to rearrange the lookups to > exploit locality, it might change the overall caching strategy. > > - As noted above, most entries are just CACHE_MISSING. So rather than > lazily looking up and caching entries, we could just prepopulate the > cache. And then you know that if an entry isn't present in the > cache, it does not exist in the tree. The downside is that you pay > to walk the all of HEAD^{tree}, even if you only have a few lookups > to do. That's a good tradeoff for pack-objects (which usually ends > up looking for every path anyway), but not for "git diff" (where you > only care about a few changed paths. > > - the cache here is static-local in the function. It should probably > at least be predicated on the tree_oid, and maybe attached to the > repository object? I think having one per repository at a time would > be fine (generally the tree_oid is set once per process, so it's not > like you're switching between multiple options). > >I've cc'd John as the author of 2386535511. But really, that was just enabling by >default the attr-tree code added by 47cfc9bd7d (attr: add flag `--source` to work >with tree-ish, 2023-01-14). Although in that original context (git check-attr) the >lack of caching would be much less important. Although this is an unbaked thought... I am not sure this will help, but I have been considering this since the trie data structure was introduced. Should we consider moving to a sparse trie with a lazy load approach? There is no current indication in the trie structure in path.c that a node is incomplete (not as far as I can tell anyway), so the assumption is the trie must be fully populated. Loading tries are generally the same performance as searching - Order(1) - given there are limited numbers of splits compared to a balanced or red-black tree; so a lazy load would not significantly change the trie load time. Each lookup might expand the node population but could cut down times where parts of the trie are ignored. Although this would be a non-trivial change, knowing that the trie is incomplete in a sparse situation might help here. --Randall
