[NTLUG:Discuss] pthreads on Linux

[Peter A. Koren]

I just wrote a threaded version (pthreads) of a C program which has one
main thread and typically 50 to 100 child threads. I have an ABIT BP-6
dual processor motherboard and am running Mandrake 8.0 with SMP kernel
version 2.4.3. I was hoping to get a speed up. No such luck.

The threaded version does use both processors at high utilization rates,
as confirmed by "top." The algorithm, BTW, is a differential evolution
code, which is a clever cross between simulated annealing and genetic
algorithms. On "toy" problems, with little computation per thread, the
non threaded version runs about an order of magnitude faster, indicating
much more overhead for threading. But even when I add a very long
running delay loop for each thread, the threaded version takes about 50%
more time.

I added more memory 128MB => 256 MB. It did nothing. I changed from
using a join approach to a detached thread model with semaphores to
synchronize threads, and it helped take 20 seconds off the threaded
benchmarks, for both the "toy" threads and the longer running threads
with built in delay loops. But I just can't find a way to make threads
pay off.

I thought that threads under Linux were quite efficient with two
processors, as that is what is widely claimed. Do any of you have any
relevant experience? Have you seen gains using threads on
multi-processor systems?

Pete Koren