git/list[1] front-page[2] threads[3] people[4] search[5] about
 

[RFC] Use a 16-tree instead of a 256-tree for storing notes

From
Johan Herland <johan@herland.net>
Date
Aug 26, 2009, 10:31 UTC
Message-ID
<200908261231.01616.johan@herland.net>
In-Reply-To
<200908010436.57480.johan@herland.net>

The 256-tree structure is considerably faster than storing all entries in a hash_map. Also, the memory consumption of the 256-tree structure is lower than the hash_map, provided that you're only loading a few notes from a "properly fanned-out" notes tree (i.e. 100000 notes in a 2/2/36 structure). However, in the worst case (loading all 100000 notes), the memory usage of the 256-tree structure (62.64 MB) is significantly worse than the hash_map approach (10.25 MB).

This patch modifies the 256-tree structure into a 16-tree structure. This significantly improves the memory situation. The result uses less memory than both the 256-tree structure, and the hash_map approach, with a worst case usage of 8.54 MB. Additionally, it seems to slightly improve the runtime performance as well (probably because of the improved memory usage).

In conclusion, this is faster and smaller than all the previous drafts.

$ ./test_performance.sh Timing 100 reps of 'git log -n 10 refs/heads/master >/dev/null' at fanout level 0... 15.05user 1.44system 0:16.59elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+782490minor)pagefaults 0swaps

Timing 100 reps of 'git log -n 10 refs/heads/master >/dev/null' at fanout level 1... 0.68user 0.17system 0:00.87elapsed 98%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+99585minor)pagefaults 0swaps

Timing 100 reps of 'git log -n 10 refs/heads/master >/dev/null' at fanout level 2... 0.41user 0.17system 0:00.61elapsed 97%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (0major+94084minor)pagefaults 0swaps

Signed-off-by: Johan Herland <johan@herland.net>
---
Hi,
This patch goes on top of the 256-tree RFC I sent earlier.

If nobody suggests further improvements, this patch will be included in the next iteration of the jh/notes topic.

Have fun! :)
...Johan
 notes.c |   26 ++++++++++++++------------
 1 files changed, 14 insertions(+), 12 deletions(-)
diff --git a/notes.c b/notes.c
index 9282b16..32b1e01 100644
--- a/notes.c
+++ b/notes.c
@@ -7,9 +7,9 @@
 #include "tree-walk.h"
 
 /*
- * Use a non-balancing simple 256-tree structure with struct int_node as
+ * Use a non-balancing simple 16-tree structure with struct int_node as
  * internal nodes, and struct leaf_node as leaf nodes. Each int_node has a
- * 256-array of pointers to its children
+ * 16-array of pointers to its children
  * The bottom 2 bits of each pointer is used to identify the pointer type
  * - ptr & 3 == 0 - NULL pointer, assert(ptr == NULL)
  * - ptr & 3 == 1 - pointer to next internal node - cast to struct int_node *
@@ -21,18 +21,18 @@
  *
  * To add a leaf_node:
  * 1. Start at the root node, with n = 0
- * 2. Use the nth byte of the key as an index into a:
- *    - If NULL, store the tweaked pointer directly into a[n]
- *    - If an int_node, recurse into that node and increment n
- *    - If a leaf_node:
+ * 2. Use the nth nibble of the key as an index into a:
+ *    - If a[n] is NULL, store the tweaked pointer directly into a[n]
+ *    - If a[n] is an int_node, recurse into that node and increment n
+ *    - If a[n] is a leaf_node:
  *      1. Check if they're equal, and handle that (abort? overwrite?)
  *      2. Create a new int_node, and store both leaf_nodes there
  *      3. Store the new int_node into a[n].
  *
  * To find a leaf_node:
  * 1. Start at the root node, with n = 0
- * 2. Use the nth byte of the key as an index into a:
- *    - If an int_node, recurse into that node and increment n
+ * 2. Use the nth nibble of the key as an index into a:
+ *    - If a[n] is an int_node, recurse into that node and increment n
  *    - If a leaf_node with matching key, return leaf_node (assert note entry)
  *    - If a matching subtree entry, unpack that subtree entry (and remove it);
  *      restart search at the current level.
@@ -42,7 +42,7 @@
  *        subtree entry (and remove it); restart search at the current level.
  */
 struct int_node {
-	void *a[256];
+	void *a[16];
 };
 
 /*
@@ -79,11 +79,13 @@ static int initialized;
 static void load_subtree(struct leaf_node *subtree, struct int_node *node,
 		unsigned int n);
 
+#define GET_NIBBLE(n, sha1) (((sha1[n >> 1]) >> ((n & 0x01) << 2)) & 0x0f)
+
 static struct leaf_node *note_tree_find(struct int_node *tree, unsigned char n,
 		const unsigned char *key_sha1)
 {
 	struct leaf_node *l;
-	unsigned char i = key_sha1[n];
+	unsigned char i = GET_NIBBLE(n, key_sha1);
 	void *p = tree->a[i];
 
 	switch(GET_PTR_TYPE(p)) {
@@ -138,7 +140,7 @@ static int note_tree_insert(struct int_node *tree, unsigned char n,
 	struct int_node *new_node;
 	const struct leaf_node *l;
 	int ret;
-	unsigned char i = entry->key_sha1[n];
+	unsigned char i = GET_NIBBLE(n, entry->key_sha1);
 	void *p = tree->a[i];
 	assert(GET_PTR_TYPE(entry) == PTR_TYPE_NULL);
 	switch(GET_PTR_TYPE(p)) {
@@ -211,7 +213,7 @@ static void load_subtree(struct leaf_node *subtree, struct int_node *node,
 		     sha1_to_hex(subtree->val_sha1));
 
 	prefix_len = subtree->key_sha1[19];
-	assert(prefix_len >= n);
+	assert(prefix_len * 2 >= n);
 	memcpy(commit_sha1, subtree->key_sha1, prefix_len);
 	while (tree_entry(&desc, &entry)) {
 		int len = get_sha1_hex_segment(entry.path, strlen(entry.path),
-- 
1.6.4.304.g1365c.dirty
Previous: Johan HerlandNext: Alex Riesen
Message 19 of 30 in “[PATCHv3 0/8] RESEND: git notes”
  1. Johan HerlandJul 29, 2009
  2. 1/8 Introduce commit notesJohan Herland, Jul 29, 2009
  3. Alex RiesenJul 29, 2009
  4. Johannes SchindelinJul 29, 2009
  5. Johan HerlandJul 30, 2009
  6. Johannes SchindelinJul 30, 2009
  7. Johan HerlandJul 30, 2009
  8. 2/8 Add a script to edit/inspect notesJohan Herland, Jul 29, 2009
  9. 3/8 Speed up git notes lookupJohan Herland, Jul 29, 2009
  10. 4/8 Add an expensive test for git-notesJohan Herland, Jul 29, 2009
  11. 5/8 Teach "-m <msg>" and "-F <file>" to "git notes edit"Johan Herland, Jul 29, 2009
  12. Thomas RastJul 29, 2009
  13. Johan HerlandJul 30, 2009
  14. 6/8 First draft of notes tree parser with support for fanout subtreesJohan Herland, Jul 29, 2009
  15. Johannes SchindelinJul 29, 2009
  16. Testing performance of the notes lookup code (Was: [PATCHv3 6/8] First draft of notes tree parser with support for fanout subtrees)Johan Herland, Jul 30, 2009
  17. [RFC] First draft of 256-tree structure for storing notesJohan Herland, Aug 1, 2009
  18. [RFC] Store subtree entries in the same hash map as the note entriesJohan Herland, Aug 13, 2009
  19. [RFC] Use a 16-tree instead of a 256-tree for storing notesJohan Herland, Aug 26, 2009
  20. Alex RiesenAug 26, 2009
  21. Johan HerlandAug 26, 2009
  22. Alex RiesenAug 26, 2009
  23. Andreas EricssonAug 26, 2009
  24. Johan HerlandAug 26, 2009
  25. Johannes SchindelinAug 27, 2009
  26. 7/8 fast-import: Add support for importing commit notesJohan Herland, Jul 29, 2009
  27. Shawn O. PearceJul 29, 2009
  28. Johan HerlandJul 30, 2009
  29. 8/8 t3302-notes-index-expensive: Speed up create_repo()Johan Herland, Jul 29, 2009
  30. Johannes SchindelinJul 29, 2009

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.