This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Performance testing shared vs. static libs
- From: Scott Bronson <bronson at rinspin dot com>
- To: gcc at gcc dot gnu dot org
- Date: Wed, 30 Jun 2004 02:58:51 -0700
- Subject: Performance testing shared vs. static libs
I wrote these tests to see how much slower it is to call into a shared
library than it is to call into a static library. My intuition would
say that shared is slower, of course, but by how much?
These timings are calling a routine that simply increments an integer
(it's only 8 instructions total), so the differences shown here are
pretty much worst-case. For example, if my test says that calling a
shared library is 10% slower, but your routine is 80 instructions long
(10x larger than mine), then for you, calling a shared library routine
would only be a 1% performance hit.
The tests shown below were run on Linux 2.6, gcc 3.3.3 on an Athlon
1800. You can retrieve the test code from funcptrs-0.2.tar.gz on:
http://www.rinspin.com/bronson/code/gcc/
COMPILED RESULTS
Shared vs. static
Calling a shared library routine is 1-20% slower, with the typical
performance hit somewhere around 15%:
direct to shared library:
-O0=5%, -01=20%, -O2=20%, -O3=1%, -Os=10% slower
indirect to shared library:
-O0=10%, -01=15%, -O2=15%, -O3=15%, -Os=10% slower
PIC (position-independent code) vs. position dependent code:
-fPIC (position independent code) doesn't affect the speed of the direct
function call at all (except for -Os, where it's 5% slower). However,
when calling via function pointer, -fPIC causes a 0 to 30% peformance
hit (-O0=0%, -O1=10%, -O2=20%, -O3=30%, -Os=15% slower) over
position-dependent code.
Static linking vs. compiling directly:
As you would expect, statically linking to a routine in a library
usually provides exactly the same performance as directly compiling the
routine into your program. There are some exceptions, however:
-O1 is 5% slower and -Os is 5% faster. This seems really weird to me.
Why would it be any different at all?
CONCLUSION
Yes, calling a shared library routine is slower. But not much. For
trivial functions, it might be 30% slower worst case, 15% typical,
depending on your code and optimization level. For real-world
functions, as long as they're not used in the innermost loops, the delay
caused by calling into a shared library is negligible.
- Scott
DATA:
For -O0:
Directly calling a shared library routine is 5% slower than calling
it statically.
Indirect function call: (5% slower than direct)
Calling into a shared library is 10% slower than calling statically.
(i.e. calling a shared library function indirectly is 15% slower
than calling a static library function directly)
null: min=0.35936 max=0.37501 avg=0.36353 44.917%
direct: min=0.80486 max=0.82058 avg=0.80934 100.000%
dirshare: min=0.82967 max=0.87794 avg=0.84628 104.563%
dirstatic: min=0.80114 max=0.80580 avg=0.80432 99.379%
dirpicstatic: min=0.80263 max=0.82554 avg=0.80936 100.002%
indirect: min=0.83391 max=0.85256 avg=0.84431 104.320%
indirshare: min=0.91188 max=0.95370 avg=0.93296 115.273%
indirstatic: min=0.83172 max=0.86983 avg=0.84892 104.889%
indirpicstatic: min=0.83234 max=0.87370 avg=0.84546 104.462%
For -O1:
calling a shared library routine is 20% slower than calling
it statically (for both direct and indirect).
Directly calling a static library routine is 5% slower than calling
a routine that has been directly compiled in (?!). This is
reproducible.
For some reason, PIC causes a 10% performance hit in the indirect
call, but not in the direct call!
null: min=0.24641 max=0.25326 avg=0.24962 48.134%
direct: min=0.51562 max=0.51982 avg=0.51859 100.000%
dirshare: min=0.62994 max=0.63484 avg=0.63222 121.911%
dirstatic: min=0.54484 max=0.54833 avg=0.54584 105.254%
dirpicstatic: min=0.54667 max=0.55310 avg=0.55051 106.156%
indirect: min=0.53638 max=0.55393 avg=0.54647 105.377%
indirshare: min=0.63173 max=0.64108 avg=0.63423 122.299%
indirstatic: min=0.51710 max=0.52236 avg=0.51930 100.137%
indirpicstatic: min=0.57687 max=0.57940 avg=0.57802 111.460%
For -O2:
calling a shared library routine is 20% slower than calling
it statically (for both direct and indirect).
Directly calling a static library routine is 5% slower than calling
a routine that has been directly compiled in (?!). This is
reproducible.
For some reason, PIC causes a 20% performance hit in the indirect
call, but not in the direct call!
null: min=0.25051 max=0.25293 avg=0.25157 48.603%
direct: min=0.51552 max=0.52036 avg=0.51761 100.000%
dirshare: min=0.63215 max=0.63767 avg=0.63372 122.432%
dirstatic: min=0.51592 max=0.51873 avg=0.51756 99.991%
dirpicstatic: min=0.51541 max=0.52058 avg=0.51775 100.027%
indirect: min=0.51572 max=0.52056 avg=0.51831 100.136%
indirshare: min=0.60469 max=0.60900 avg=0.60630 117.135%
indirstatic: min=0.51512 max=0.51920 avg=0.51735 99.950%
indirpicstatic: min=0.63046 max=0.64098 avg=0.63390 122.466%
For -O3:
It doesn't matter if youre callind a shared library, static
library, or your own code, or -fPIC or not. All direct calls
are very close to each other (about 1%).
We do see that indirectly calling your own code or a static
library goes 10% faster than direct (as found in the previous
battery of tests).
Indirectly calling a shared library takes 15% longer than
indirectly calling a static library (and 5% longer than directly
calling either static or shared).
PIC continues to cause a 20% performance hit for indirect.
null: min=0.24947 max=0.25187 avg=0.25082 43.595%
direct: min=0.57265 max=0.57676 avg=0.57534 100.000%
dirshare: min=0.58120 max=0.58617 avg=0.58403 101.509%
dirstatic: min=0.57705 max=0.59201 avg=0.58076 100.942%
dirpicstatic: min=0.57274 max=0.57839 avg=0.57535 100.001%
indirect: min=0.51557 max=0.51962 avg=0.51780 89.999%
indirshare: min=0.60617 max=0.60948 avg=0.60813 105.699%
indirstatic: min=0.51572 max=0.51917 avg=0.51813 90.056%
indirpicstatic: min=0.62970 max=0.63533 avg=0.63271 109.970%
For -Os:
Calling a shared library takes 10% longer than static (both direct and
indirect)
Calling a static library is 5% faster than calling direct code?!?!
-fPIC causes a 5% performance hit direct and a 15% performance hit
indirect.
null: min=0.25008 max=0.25290 avg=0.25153 43.712%
direct: min=0.57253 max=0.57886 avg=0.57543 100.000%
dirshare: min=0.62959 max=0.63425 avg=0.63178 109.793%
dirstatic: min=0.54431 max=0.54895 avg=0.54653 94.977%
dirpicstatic: min=0.57430 max=0.58078 avg=0.57699 100.270%
indirect: min=0.57590 max=0.57842 avg=0.57706 100.283%
indirshare: min=0.63282 max=0.63675 avg=0.63517 110.382%
indirstatic: min=0.57224 max=0.57681 avg=0.57524 99.966%
indirpicstatic: min=0.65924 max=0.66185 avg=0.66071 114.819%