condition wakeup behaviour

I am running a comparision test with one driver thread and lots of
worker threads; and seeing some surprising behaviour. When I run with
old style linux pthreads, the test scales badly as the number of threads
increase (this is expected); the same behaviour happens on SMP systems
and with different linux kernel schedulers.  When I run the test with
NGPT the test scales well on a 1 cpu system, but seems to pause when
running on an 8-way system. 

I am running a patched 2.4.18 kernel with NGPT 1.2.2

Perhaps the problem is in the pthread_cond_signal (pth_cond_notify)
since it just makes the condition change state then starts yielding to
all the threads.  Wouldn't it be better to keep some state in the
condition variable that points to the thread on the "top of the waiting
queue" and yield directly to that thread?