Re: [PATCH 4/4] Revert "local.conf.sample: Disable image-prelink by default"

[Mark Hatle <mark.hatle@windriver.com>]

On 11/2/15 6:10 PM, Khem Raj wrote:
> On Mon, Nov 2, 2015 at 4:07 PM, Mark Hatle <mark.hatle@windriver.com> wrote:
>> This reverts commit 6dd28030f323d7106a02ec54ce4e249561ab0836.
>>
>> Prelink now works properly again.  Re-enable the functionality.
>>
>> Signed-off-by: Mark Hatle <mark.hatle@windriver.com>
>> ---
>>  meta/conf/local.conf.sample | 4 +---
>>  1 file changed, 1 insertion(+), 3 deletions(-)
>>
>> diff --git a/meta/conf/local.conf.sample b/meta/conf/local.conf.sample
>> index 9d1480e..3ae24ab 100644
>> --- a/meta/conf/local.conf.sample
>> +++ b/meta/conf/local.conf.sample
>> @@ -128,9 +128,7 @@ EXTRA_IMAGE_FEATURES = "debug-tweaks"
>>  #   - 'image-swab' to perform host system intrusion detection
>>  # NOTE: if listing mklibs & prelink both, then make sure mklibs is before prelink
>>  # NOTE: mklibs also needs to be explicitly enabled for a given image, see local.conf.extended
>> -# NOTE: image-prelink is currently broken due to problems with the prelinker.  It is advised
>> -# that you do NOT run the prelinker at this time.
>> -USER_CLASSES ?= "buildstats image-mklibs"
>> +USER_CLASSES ?= "buildstats image-mklibs image-prelink"
> 
> With all the discussions around benefits of it recently, I think its
> better disabled IMO,
> 

I disagree.  I've got rough estimate figures for prelink improvement to the system.

core-image-sato w/ systemd enabled:

MIPS64 - n64 - 1.5MB of ram saved
MIPS64 - n32 - 500KB of ram saved
MIPS32 - n32 - 300KB of ram saved

ARM - 560k of ram saved

PPC - 380k of ram saved

IA32 - forgot to write it down, was around 500k.


Testing also showed sysvinit saves even more memory due to COW pages then
systemd.  (I was a bit surprised by that.)  My testing consisted of building a
filesystem and booting it.  Waiting for everything to 'finish' and then
comparing the output of the 'free' command in KB.


As for startup time savings, I don't have specific figures BUT, for a few key
cases the load time of the executable is up to 50% faster.  For instance busybox
on the qemux86-64.  Keep in mind that busybox is pretty simple and uses only a
few libraries.  The more libraries in use, the more prelink will affect load times.

The busybox binary loading and then exiting (LD_BIND_NOW=1 LD_DEBUG=statistics):

(not prelinked with LD_BIND_NOW defined to simulate worst case, all relocations
needed)

       566:
       566:     runtime linker statistics:
       566:       total startup time in dynamic loader: 182539970 cycles
       566:                 time needed for relocation: 40894259 cycles (22.4%)

       566:                      number of relocations: 443
       566:           number of relocations from cache: 3
       566:             number of relative relocations: 1200
       566:                time needed to load objects: 63004343 cycles (34.5%)

       566:
       566:     runtime linker statistics:
       566:                final number of relocations: 445
       566:     final number of relocations from cache: 3

(prelinked, same LD_BIND_NOW defined)

       524:
       524:     runtime linker statistics:
       524:       total startup time in dynamic loader: 32420607 cycles
       524:                 time needed for relocation: 2812608 cycles (8.6%)
       524:                      number of relocations: 4
       524:           number of relocations from cache: 76
       524:             number of relative relocations: 0
       524:                time needed to load objects: 15549810 cycles (47.9%)

       524:
       524:     runtime linker statistics:
       524:                final number of relocations: 6
       524:     final number of relocations from cache: 76

(both of the above were run with a hot cache, multiple times until the system
stabilized and returned results that were fairly consistent)

So the executable load time dropped to 17% (32420607/182539970) of the original
value due to being prelinked.