optimising opennmpi communication within a 10/100Mbs Ethernet connection using OTPO tool

[Alfred]

Hi
i want to improve communication performance of my cluster by optimising my openmpi, i have thought of minimising latency as way to tackle this, i have downloaded the OTPO tool. (Open Tool For Parameter Optimisation), and now want to use the tool to optmise mt openmpi within my 10/100Mbps, i think its something to do with some Specified MCA parameters which for the moment am not of aware of. could anyone pliz give me some hints to get started. i feel stuck yet i know what i want to achieve.

[Alfred]

I set up my cluster using pelicanhpc, and run the example programs, so i want to improve the computing performance of the programs by improving the timing of the programs, so thats why go into the latency issue like i explain in my other post.

[Michael Creel]

I don't know anything about this, I normally just start up the cluster and get to work. If you're able to improve timings, I'd be very interested to hear about it.

[Alfred]

well from my findings, it seems the only option to improve performance is by upgrading the ethernet switch am having of 10/100Mbps to 1000Mbps ~ a Gigabit Ethenet, which involves re-investing in the system infrastructure. it will lead to a significant performance improvement from the communication perspective.  may be increasing the sustained bandwidth from about ~ 12Mbps to ~100Mbps.
there is no need to tune the openmpi, because the default parameters have already been set for the best achievable performance. however its important that your networking cards in the PC have to support 1000Mbps.

i tried to test my cluster  with the kernel_example, changed N from 2000, TO 5000 TO 10000 TO 20000, and observed the timing on my nodes,, which is reasonable, though as N gets bigger to 10000 and above, it cant be run on the master node alone, so when i split the work using mpi across the compute nodes, i will get results but i cant get the graphs output display like when i run kernel_example at octave prompt for a smaller value of N like 2000,.and get results.

[Michael Creel]

I haven't looked at that code for a while, but it should be possible to do the computations using compute nodes, and then display the graph on the frontend.