Question

I am learning the concept of threads in c/c++. I was trying the examples for
the dot product of two vectors cited in Pthreads Overview.

• Serial version

• Pthreads version

I ran both the serial and the thread version of the code and I found that the
serial version was faster than the thread version. I thought it should be the
opposite.

I am running on a single CPU.

Answer 1

The code that link to has a few issues that you need to keep in mind:

1. The serial version is doing an inner product of two length-100 vectors. The parallel version is doing an inner product of two length-400 vectors. You can see that in the array allocation (e.g., a = (double*) malloc (NUMTHRDS*VECLEN*sizeof(double)); — NUMTHRDS is set to four and VECLEN 100 in the code). Therefore, the parallel program is doing four times the amount of work, but with four threads, so the naive assumption is that the serial and parallel programs will have the same run time.
2. The parallel code is demonstrating a mutex for thread synchronization. This may be creating barriers in the code as it runs.
3. The code is using four threads. If your CPU does not have four or more threads, then you cannot expect it to scale.
4. There is an overhead in creating the threads, and with such a small problem it is probably a significant factor.