| 1 |
Mike Gilbert <floppym@g.o> wrote: |
| 2 |
|
| 3 |
> > Well, bosh has been tested to work as /bin/sh on Gentoo. |
| 4 |
> > BTW: On Solaris, bosh is faster than dash (because Solaris has a fully working |
| 5 |
> > vfork()). On Linux bosh is "only" of the same speed as dash since vfork() on |
| 6 |
> > Linux does not borrow the parents address space description but copies it. |
| 7 |
> |
| 8 |
> Is that also true of clone(CLONE_VM|CLONE_VFORK)? Recent versions of |
| 9 |
> glibc use this to implement the posix_spawn() function. |
| 10 |
|
| 11 |
I guess... from reading the man page. |
| 12 |
|
| 13 |
There is a simple test to verify my clain for vfork(): |
| 14 |
|
| 15 |
bosh may easily be configured for various behavior. If you like a bosh that |
| 16 |
does not use vfork() just call: |
| 17 |
|
| 18 |
cd sh |
| 19 |
smake clean |
| 20 |
smake COPTX=-DNO_VFORK |
| 21 |
|
| 22 |
If you run tests like: |
| 23 |
|
| 24 |
CONFIG_SHELL=$shell $shell ./configure |
| 25 |
|
| 26 |
in the directory "autoconf" from the schilytools: |
| 27 |
|
| 28 |
http://sourceforge.net/projects/schilytools/files/ |
| 29 |
|
| 30 |
(first run "rm *cache"), there is a significant performance win on Solaris |
| 31 |
for the version that uses vfork() - 20% less kernel CPU time. |
| 32 |
|
| 33 |
On Linux, there is a too little win for the same comparison. |
| 34 |
|
| 35 |
On Solaris, a vfork() call (check with truss -c) is takes only 1/3 of the time |
| 36 |
of a fork(), on Linux the win is only 30%. The total performance win for bosh |
| 37 |
executing configure using vfork() on Solaris is 10-15%, on Linux the total win |
| 38 |
is only 3%. |
| 39 |
|
| 40 |
Note that Linux copied the Solaris concept to implement a normal fork() with |
| 41 |
copy on write mmaps to the parent. So the only difference between fork() and |
| 42 |
vfork() is the handling of the address space description for the processes |
| 43 |
in the kernel. |
| 44 |
|
| 45 |
I would be happy, if there could be a way to get the same vfork() performance win |
| 46 |
on Linux as seen on Solaris. |
| 47 |
|
| 48 |
BTW: I forgot to mention that the unit tests for bosh are used this way: |
| 49 |
|
| 50 |
- first compile schilytools by running smake from top level. |
| 51 |
|
| 52 |
- Then chdir to "sh" and call "smake tests", see the global code |
| 53 |
in $SRCROOT/tests and the specific code in sh/tests |
| 54 |
|
| 55 |
Jörg |
| 56 |
|
| 57 |
-- |
| 58 |
EMail:joerg@××××××.net (home) Jörg Schilling D-13353 Berlin |
| 59 |
joerg.schilling@××××××××××××××××.de (work) Blog: http://schily.blogspot.com/ |
| 60 |
URL: http://cdrecord.org/private/ http://sf.net/projects/schilytools/files/' |
Archives