I don't fully understand your question but one thing is that the i7 doesn't have eight physical cores. It has four cores and multithreading (am i right?). So using more than four cores can under certain circumstances affect speed negatively, no matter if you're using MPI or OpenMP.

Dear Niko Komin,

Thanks for your answer...it adds a new angle to my approach towards performance evaluation of parallel CFD codes. The issue here is about understanding the differences between the terms "processor", "core" and "thread" for i7. Actually this link to a question on stackoverflow sheds some more light in support of your comments...

https://stackoverflow.com/questions/10403201/8-logical-threads-at-4-cores-will-at-a-maximum-run-4-times-faster-in-parallel

I think I may need to rephrase my question as "How much is the maximum speedup obtainable from a single node i7 machine?" :)

Thanks for the help...I really appreciate it.

Best regards.

Shaswat Saincher

I found one more link that I'd like to share...

https://www.pugetsystems.com/labs/hpc/Hyper-Threading-may-be-Killing-your-Parallel-Performance-578/

Best regards.

Shaswat

Exacty. I don't know  if 'disabling' hyper threading on your personal machine makes any sense. I  just tell my program to use only four parallel executions.

Dear Niko Komin,

I did try disabling hyperthreading from the BIOS. Here's what I observed:

ψ(Np=4) hyperthreading enabled-------> 3.141

ψ(Np=4) hyperthreading disabled-------> 3.433

Hence, as explained by Dr. Kinghorn, disabling hyperthreading on i7 does lead to an improvement in parallel performance. I guess the improvement is not very dramatic because the code is already being executed in Linux. This finding is also in concurrence with the observations of Dr. Kinghorn.

I again thank you for prompting me to think in this direction, I greatly appreciate your help.

Best regards.

Shaswat Saincher

Hyperthreading is a technique that tries to maximize the number of instructions executed in a certain amount of time. Even though it looks like there are more cores logically, the system has the physical cores only. it means that if you use all the logical cores, you cannot expect the linear speedup. It is known that, on a single socket, a hyperthrraded core can provide about 30% performance improvement. If the number of socket is doubled, the performance improvement is halved.

In general, a well designed MPI code should run fine on a set of ranks that are all in the same SMP domain. It can sometimes even outperform a comparable code using OpenMP. So, normally it would not be worth rewriting the code to switch from using MPI to OpenMP. I agree with earlier comments that it is better to limit to one MPI rank/core and avoid the hyperthreads.

Dear Sunwoo Lee and Bill Long,

I am grateful for your comments.

In our case, what's happening is that the speedup decreases after we exceed the number of physical cores on our i7 system (SMP architecture). As advised in the comments, I will try avoiding hyper-threading either by using Np

Greetings everyone.

I just came across some papers discussing the impact of hyper-threading in HPC environments....I think they are worth sharing:

Article An Empirical Study of Hyper-Threading in High Performance Co...

https://pdfs.semanticscholar.org/df61/e878fa9052d85abe8fb341ee6d42b4f4c6c7.pdf

Best regards.

Shaswat Saincher