On Mon, Sep 30, 2013 at 10:38, John Carr wrote:
Here is a diff relative of OpenBSD 5.3 to avoid taking a lock in
pthread_getspecific in the common case where storage for the key is
already allocated.
I have only tested this patch for my use case where a single key is
created once and used
Here is a diff relative of OpenBSD 5.3 to avoid taking a lock in
pthread_getspecific in the common case where storage for the key is
already allocated.
I have only tested this patch for my use case where a single key is
created once and used many times. In that case, the bottleneck I
observed is
I got an email saying unified diff is preferred, so here's a resend
in that format.
Index: rthread_tls.c
===
RCS file: /cvs/src/lib/librthread/rthread_tls.c,v
retrieving revision 1.13
diff -u -r1.13 rthread_tls.c
--- rthread_tls.c 6
I attached a program showing the slowdown. It spawns threads that call
pthread_getspecific in a tight loop. On Linux the wall clock time is
essentially constant for number of threads up to number of processors.
On OpenBSD 5.3 wall clock time increases approximately linearly with
number of
I think _spinlock is suboptimal, although that's not the real problem
as far as my code is concerned. Spinlock is a loop:
while (_atomic_lock(lock-ticket)) _sched_yield();
This causes a system call every time the lock is held by another thread.
In many cases the spinlock protects a simple
We haven't done a lot of work to optimize performance except in
response to specific issues. Sounds like you found one. Would you
mind providing a test case? I just want to make sure we fix the
right thing.
On Wed, Sep 25, 2013 at 13:12, John Carr wrote:
Problem in a nutshell:
