open TELEMAC-MASCARET The mathematically superior suite of solvers

Hello all,

I carried out some benchmarks on my workstation PC in order to compare speed between intel and gfortran compilers. The benchmark revealed some interesting results regarding the the efficiency of my intel i7 CPU to run parallel jobs in multiple cores.

The benchmark test case is malpasset-large.

The PC configuration is the following:
Intel Core i7 860@ 2.80GHz (4 cores)
16Gb DDR3 Dual channel memory @667 MHz
Windows 8 x64

Compiler versions:
Intel fortran v13.0.0.89
gfortran v3.8.0

First of all, I run the case using 1,2,3 and 4 cores. The 1 core run was done using the parallel module, not the scalar.



It becomes obvious that:

• Intel fortran is a bit faster that gfortran (~14% using 2 cores)
• The benefit of using more that 2 cores is minimal

Having seen that, i tried running 2 simulations at a time using 1 core and 2 cores. The results are disappointing:




It took significant more time to complete each simulation, even when each simulation was running on a single core. It is thus preferable to run the computations the one after the other and not at the same time.

I hope this provides some insight on calculation speed aspects.

Costas

Hello,

Thank you for this insight. Please note that access to disk is your limiting factor here. I am guessing you do not have parallel access to disk when saving results.

To get a better measure, I would recommend you run over longer periods, with virtually no results printed to disk.

Thank you,
Sébastien.

Hello,

I believe at that time I was working in my workstation which has parallel disk access. Now that I am using Windows HPC, I run the computations over the network and the only time I find that network is the bottleneck, is when I split a computation between 2 or more nodes.

In order to eliminate disk limitations, I have setin the malpasset cas file (which I assumed it reduces disk access) and run the benchmark again, both over the network and locally. The results were almost identical.

I have updated the previous chart with my latest runs. The poor scaling between 2-core and 4-core runs remains. A slight performance improvement can be observed, which could be down to the latest version of gcc (4.8.3 vs 4.8.0) or MSMPI (previous results were obtained using MPICH2)



Regards,
Costas