Saturday, July 10, 2004
Is profiling even viable now?
Mark brought up (in email) an interesting optimization technique using GCC 3:
I came across an interesting optimization that is GCC specific but quite clever.
In lots of places in the Linux kernel you will see something like:
p = get_some_object(); if (unlikely(p == NULL)) { kill_random_process(); return (ESOMETHING); } do_stuff(p);The conditional is clearly an error path and as such means it is rarely taken. This is actually a macro defined like this:
#define unlikely(b) __builtin_expect(b, 0)On newer versions of GCC this tells the compiler to expect the condition not to be taken. You could also tell the compiler that the branch is likely to be taken:
#define likely(b) __builtin_expect(b, 1)So how does this help GCC anyhow? Well, on some architectures (PowerPC) there is actually a bit in the branch instruction to tell the CPU's speculative execution unit if the branch is likely to be taken. On other architectures it avoids conditional branches to make the “fast path” branch free (with
-freorder-blocks).
I was curious to see if this would actually help any, so I found a
machine that had GCC 3 installed (swift), compiled a version of
mod_blog with profiling
information, ran it, found a function that looked good to speed up, added
some calls to __builtin_expect(), reran the code and got a
rather encouragine interesting result.
I then reran the code, and got a completely different result.
In fact, each time I run the code, the profiling information I get is nearly useless—well, to a degree. For instance one run:
| % time | cumulative seconds | self seconds | calls | self ms/call | total ms/call | name |
|---|---|---|---|---|---|---|
| 100.00 | 0.01 | 0.01 | 119529 | 0.00 | 0.00 | line_ioreq |
| 0.00 | 0.01 | 0.00 | 141779 | 0.00 | 0.00 | BufferIOCtl |
| 0.00 | 0.01 | 0.00 | 60991 | 0.00 | 0.00 | line_readchar |
| 0.00 | 0.01 | 0.00 | 59747 | 0.00 | 0.00 | ht_readchar |
Then another run:
| % time | cumulative seconds | self seconds | calls | self ms/call | total ms/call | name |
|---|---|---|---|---|---|---|
| 33.33 | 0.01 | 0.01 | 119529 | 0.00 | 0.00 | line_ioreq |
| 33.33 | 0.02 | 0.01 | 60991 | 0.00 | 0.00 | line_readchar |
| 33.33 | 0.03 | 0.01 | 21200 | 0.00 | 0.00 | ufh_write |
| 0.00 | 0.03 | 0.00 | 141779 | 0.00 | 0.00 | BufferIOCtl |
Yet another run:
| % time | cumulative seconds | self seconds | calls | self ms/call | total ms/call | name |
|---|---|---|---|---|---|---|
| 0.00 | 0.00 | 0.00 | 141779 | 0.00 | 0.00 | BufferIOCtl |
| 0.00 | 0.00 | 0.00 | 119529 | 0.00 | 0.00 | line_ioreq |
| 0.00 | 0.00 | 0.00 | 60991 | 0.00 | 0.00 | line_readchar |
| 0.00 | 0.00 | 0.00 | 59747 | 0.00 | 0.00 | ht_readchar |
And still another one:
| % time | cumulative seconds | self seconds | calls | self ms/call | total ms/call | name |
|---|---|---|---|---|---|---|
| 50.00 | 0.01 | 0.01 | 60991 | 0.00 | 0.00 | line_readchar |
| 50.00 | 0.02 | 0.01 | 1990 | 0.01 | 0.01 | HtmlParseNext |
| 0.00 | 0.02 | 0.00 | 141779 | 0.00 | 0.00 | BufferIOCtl |
| 0.00 | 0.02 | 0.00 | 119529 | 0.00 | 0.00 | line_ioreq |
Like I said, nearly useless. Sure, there are the usual suspects, like
BufferIOCtl() and line_ioreq(), but it's
impossible to say what improvements I'm getting by doing this. And by
today's standards, swift isn't a fast machine being only
(only!) a 1.3GHz Pentium III
with half a gig of RAM. I
could only imagine the impossibility of profiling under a faster machine, or
even imagining what could be profiled under a faster machine.
I have to wonder what the Linux guys are smoking to even think, in the
grand scheme of things, if __builtin_expect() will even improve
things all that much.
Unless they have access to better profiling mechanics than I do.
Looks like I might have to find a slower machine to get a better feel for how to improve the speed of the program.
Profiling is still viable, if you run the program long enough
So last night I said, “I might have to find a slower machine to get a better feel for how to improve the speed of [mod_blog].” Well, I found one other way—increase the amount of work the program does.
mod_blog, or at least the program I was working with last
night, bp (for “build page”) is responsible for building the
HTML pages served up
the webserver. In its default mode (which is what I was working with last
night) it generates the main index page, and the RSS file, which only
(only!) generates about 100,000 bytes of output and happened too
quickly to get any meaningful data out of profiling the program. But it can
also generate pages with
hundreds of entries given the right options, so by having it generate a
page with every entry in 2000 through 2003 inclusive (which
generates a page that is 3,352,028 bytes in size) I was able to get
meaningful profiling information:
| % time | cumulative seconds | self seconds | calls | self ms/call | total ms/call | name |
|---|---|---|---|---|---|---|
| 29.73 | 0.22 | 0.22 | 5817769 | 0.00 | 0.00 | line_ioreq |
| 14.86 | 0.33 | 0.11 | 6650539 | 0.00 | 0.00 | BufferIOCtl |
| 10.81 | 0.41 | 0.08 | 2908866 | 0.00 | 0.00 | ht_readchar |
| 9.46 | 0.48 | 0.07 | 790760 | 0.00 | 0.00 | ufh_write |
| 5.41 | 0.52 | 0.04 | 2908398 | 0.00 | 0.00 | line_readchar |
Yes, with a run time of nearly 3 seconds (2.992) I was able to generate
consistant, meaningful profiling information (meaning, we went from calling
BufferIOCtl() 141,779 times to 6,650,539 times) which was good
enough to see if the GCC optimzation of using -freorder-blocks
with __builtin_expect() would help any, at least on the Intel
x86 line.
Well, after running a dozen tests, I can say that it does help
on the Intel platform, but at best, using __builtin_expect()
with -freorder-blocks will only give you a few percent boost in
speed. As in, single digit percentage boost.
In certain cases.
Certainly, when I did tests generating the output to
/dev/null you could see the boost, but in running tests
generating an actual file (as opposed to just tossing the data in the bit
bucket), it's not quite so clear cut (and for each test, I ran it five
times, taking the timings from the fifth run—this to help smooth out any
system caching effects). The best improvement came when using -O3
-march=pentium3
-fomit-frame-pointer to compile the program, but it was still hardly
noticable from a user perspective (maybe about a tenth or two tenths of a
second).
Mark is expecting __builtin_expect() to have better impact
on systems where the CPU
will use the hint.
