Hello,
At present we are using "sys v semaphores"(USE_SYSV_SEM) for locking.
But we see a some delay while acquiring the lock to modify the usrloc's location table.
We are in a thought to use "posix semaphores" (USE_POSIX_SEM), will this improve any performance for usrloc location table.
Thoughts and Ideas will be appreciated.
Hi, it probably highly depends on your OS and kernel.
I did performance testing on Debian linux with kernel 2.6.8 and it seems to me, that all locking methods ("fast locking", sysV, pthreads, possix) were approx. the same for "short" operations like transaction handling.
I did another testing on Centos based distro, with kernel 2.6.9 for "long" operations like usrloc with slow DB and there most of the locking methods were the same (sys V, pthreads, possix) but the "fast locking" was significantly worse!
So ... I know that I have to run these tests on the same machines etc, but for the first look it seems to me that whatever locking method than "fast locking" will be OK. But this can be valid for Linux and quite new kernel only; I don't absolutely now what about other OS.
Vaclav
On Thu, Feb 01, 2007 at 04:05:56PM +0530, Ramu Yadav wrote:
Hello,
At present we are using "sys v semaphores"(USE_SYSV_SEM) for locking.
But we see a some delay while acquiring the lock to modify the usrloc's location table.
We are in a thought to use "posix semaphores" (USE_POSIX_SEM), will this improve any performance for usrloc location table.
Thoughts and Ideas will be appreciated.
-- Ramu Yadav
Serusers mailing list Serusers@lists.iptel.org http://lists.iptel.org/mailman/listinfo/serusers
[no crosspost to serdev as I think this is a users' question]
Vaclav Kubart wrote:
it probably highly depends on your OS and kernel.
True.
We had some trouble using posix semaphores on ia32 with a Debian linux and kernel 2.6.18, while the very same seemed to work on an ia64 architecture (not sure which kernel version, but might have been a 2.6.18 as well).
To me,it seems that the performance penalty from fast locking is very small, so this method also qualifies for use in most production environments---depending on the size of the user base, of course.
Regards, Olaf
In my case was the throughput decrease with "fast locking" quite significant (about 20%).
I will do more tests with final Ottendorf (anybody knows when it will be released?) and publish results on iptel's web.
Vaclav
On Fri, Feb 02, 2007 at 10:08:21AM +0100, Olaf Bergmann wrote:
[no crosspost to serdev as I think this is a users' question]
Vaclav Kubart wrote:
it probably highly depends on your OS and kernel.
True.
We had some trouble using posix semaphores on ia32 with a Debian linux and kernel 2.6.18, while the very same seemed to work on an ia64 architecture (not sure which kernel version, but might have been a 2.6.18 as well).
To me,it seems that the performance penalty from fast locking is very small, so this method also qualifies for use in most production environments---depending on the size of the user base, of course.
Regards, Olaf
On Feb 01, 2007 at 16:05, Ramu Yadav contactramu@gmail.com wrote:
Hello,
At present we are using "sys v semaphores"(USE_SYSV_SEM) for locking.
But we see a some delay while acquiring the lock to modify the usrloc's location table.
We are in a thought to use "posix semaphores" (USE_POSIX_SEM), will this improve any performance for usrloc location table.
Yes posix sems will imporve performance a lot. Just make sure your os supports them for locking between processes (if you're on linux it's ok). Also the default "fast_lock" work very well if you're not doing time consuming stuff inside them (like db lookups).
sys v locks are very slow.
Thoughts and Ideas will be appreciated.
Andrei
Thanks!
Yes we enabled "posix sems" and we are observing the difference.
On 2/2/07, Andrei Pelinescu-Onciul andrei@iptel.org wrote:
On Feb 01, 2007 at 16:05, Ramu Yadav contactramu@gmail.com wrote:
Hello,
At present we are using "sys v semaphores"(USE_SYSV_SEM) for locking.
But we see a some delay while acquiring the lock to modify the usrloc's location table.
We are in a thought to use "posix semaphores" (USE_POSIX_SEM), will this improve any performance for usrloc location table.
Yes posix sems will imporve performance a lot. Just make sure your os supports them for locking between processes (if you're on linux it's ok). Also the default "fast_lock" work very well if you're not doing time consuming stuff inside them (like db lookups).
sys v locks are very slow.
Thoughts and Ideas will be appreciated.
Andrei
-
Andrei Pelinescu-Onciul -
Olaf Bergmann -
Ramu Yadav -
Vaclav Kubart