[gmx-users] No improvement in scaling on introducing flow control

[himanshu khandelia]

Hi Carsten,

The benchmarks were made is 1 NIC/node, and yet the scaling is bad.
Does that mean that there is indeed network congestion ? We will try
using back to back connections soon,

-himanshu



 maybe your problem is not even flow control, but the limited network
bandwidth which is shared among 4 CPUs in your case. I also have done
benchmarks on Woodcrests
> (2.33 GHz) and was not able to scale an 80000 atom system beyond 1 node with Gbit Ethernet. Looking in more detail, the time gained by the additional 4 CPUs of a
> second node was exactly balanced by the extra communication. I used only 1 network interface for that benchmark, leaving effectively only 1/4 th of the bandwidth
> for each CPU. Using two interfaces with OpenMPI did not double the network performance on our cluster. In my tests nodes with 2 CPUs sharing one NIC were faster
> than nodes with 4 CPUs sharing two NICs. Could be on-node contention, since both interfaces probably end up on the same bus internally.
>
> Are the benchmarks made with 1 or 2 NICs/node? If they are for 1 NIC/node then there should be no network congestion for the case of 8 CPUs (=2 nodes). You could
> try a back-to-back connection between two nodes to be absolutely shure that the rest of the network (switch etc.) does not play a role. I would try that and repeat
> the benchmark for 8 CPUs. See if you get a different value.
> ##############