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

Re: [PATCH] write-tree performance problems

From
CMChris Mason <mason@suse.com>
Date
Apr 20, 2005, 19:04 UTC
Message-ID
<200504201504.59541.mason@suse.com>
In-Reply-To
<Pine.LNX.4.58.0504201040400.6467@ppc970.osdl.org>
On Wednesday 20 April 2005 13:52, Linus Torvalds wrote:
Show 7 quoted lines
> On Wed, 20 Apr 2005, Chris Mason wrote:
> > The patch below with your current tree brings my 100 patch test down to
> > 22 seconds again.
>
> If you ever have a cache_entry bigger than 16384, your code will write
> things out in the wrong order (write the new cache without flushing the
> old buffer).
Whoops
Show 5 quoted lines
> Finally, if you really want to go fast, you should really try to make your
> writes powers-of-two, ie fill up the buffer entirely rather than saying
> "if I were to overflow, flush it now". It doesn't matter that much for
> some filesystems (especially local and append-only like the patterns are
> here), but it can definitely matter for the stupid ones.

Well, the difference there should be pretty hard to see with any benchmark. But I was being lazy...new patch attached. This one gets the same perf numbers, if this is still wrong then I really need some more coffee.

-chris

--- linus.back/read-cache.c 2005-04-20 10:14:23.268310000 -0400 +++ linus/read-cache.c 2005-04-20 14:54:28.554518320 -0400

@@ -232,11 +232,13 @@
 	SHA_CTX c;
 	struct cache_header hdr;
 	int i;
+	#define BUFLEN 16384
+	static char buf[BUFLEN];
+	int len = 0;
 
 	hdr.hdr_signature = htonl(CACHE_SIGNATURE);
 	hdr.hdr_version = htonl(1);
 	hdr.hdr_entries = htonl(entries);
-
 	SHA1_Init(&c);
 	SHA1_Update(&c, &hdr, offsetof(struct cache_header, sha1));
 	for (i = 0; i < entries; i++) {
@@ -246,13 +248,37 @@
 	}
 	SHA1_Final(hdr.sha1, &c);
 
-	if (write(newfd, &hdr, sizeof(hdr)) != sizeof(hdr))
-		return -1;
-
+	/* hdr is small right now, but just
+	 * in case someone changes that...
+	 */
+	if (sizeof(hdr) < BUFLEN) {
+		memcpy(buf, &hdr, sizeof(hdr));
+		len += sizeof(hdr);
+	} else {
+		if (write(newfd, &hdr, sizeof(hdr)) != sizeof(hdr))
+			return -1;
+	}
 	for (i = 0; i < entries; i++) {
 		struct cache_entry *ce = cache[i];
 		int size = ce_size(ce);
-		if (write(newfd, ce, size) != size)
+		char *p = (char *)ce;
+		while(size > 0) {
+			int count = size;
+			if (count > BUFLEN - len)
+				count = BUFLEN - len;
+			memcpy(buf + len, p, count);
+			size -= count;
+			len += count;
+			p += count;
+			if (len == BUFLEN) {
+				if (write(newfd, buf, len) != len)
+					return -1;
+				len = 0;
+			}
+		}
+	}
+	if (len) {
+		if (write(newfd, buf, len) != len)
 			return -1;
 	}
 	return 0;
Previous: Linus TorvaldsNext: Linus Torvalds
Message 40 of 54 in “write-tree performance problems”
  1. write-tree performance problemsChris Mason, Apr 19, 2005
  2. Linus TorvaldsApr 19, 2005
  3. Chris MasonApr 19, 2005
  4. Linus TorvaldsApr 19, 2005
  5. Chris MasonApr 19, 2005
  6. Linus TorvaldsApr 19, 2005
  7. Chris MasonApr 20, 2005
  8. Linus TorvaldsApr 20, 2005
  9. Linus TorvaldsApr 20, 2005
  10. H. Peter AnvinApr 20, 2005
  11. WARNING! Object DB conversion (was Re: [PATCH] write-tree performance problems)Linus Torvalds, Apr 20, 2005
  12. Ingo MolnarApr 20, 2005
  13. Jon SeymourApr 20, 2005
  14. Martin UeckerApr 20, 2005
  15. Morten WelinderApr 20, 2005
  16. Jon SeymourApr 20, 2005
  17. C. Scott AnanianApr 20, 2005
  18. Martin UeckerApr 20, 2005
  19. C. Scott AnanianApr 20, 2005
  20. Martin UeckerApr 20, 2005
  21. Martin UeckerApr 20, 2005
  22. Blob chunking code. [First look.]C. Scott Ananian, Apr 20, 2005
  23. Blob chunking code. [Second look]C. Scott Ananian, Apr 20, 2005
  24. David WoodhouseApr 20, 2005
  25. Linus TorvaldsApr 20, 2005
  26. David WoodhouseApr 20, 2005
  27. Chris MasonApr 20, 2005
  28. C. Scott AnanianApr 20, 2005
  29. Linus TorvaldsApr 20, 2005
  30. C. Scott AnanianApr 20, 2005
  31. Linus TorvaldsApr 20, 2005
  32. Linus TorvaldsApr 20, 2005
  33. David WillmoreApr 20, 2005
  34. Linus TorvaldsApr 20, 2005
  35. Linus TorvaldsApr 20, 2005
  36. Chris MasonApr 20, 2005
  37. Linus TorvaldsApr 20, 2005
  38. Chris MasonApr 20, 2005
  39. Linus TorvaldsApr 20, 2005
  40. Chris MasonApr 20, 2005
  41. Linus TorvaldsApr 20, 2005
  42. Linus TorvaldsApr 20, 2005
  43. David S. MillerApr 20, 2005
  44. David LangApr 19, 2005
  45. Linus TorvaldsApr 19, 2005
  46. David LangApr 19, 2005
  47. Linus TorvaldsApr 19, 2005
  48. David LangApr 19, 2005
  49. Linus TorvaldsApr 19, 2005
  50. Christopher LiApr 19, 2005
  51. Olivier GalibertApr 19, 2005
  52. C. Scott AnanianApr 19, 2005
  53. Linus TorvaldsApr 20, 2005
  54. C. Scott AnanianApr 20, 2005

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.