airborne

I must admit it was somewhat of a struggle getting  up. I did manage to get some stuff sorted albeit slowly. I was very pleased to figure out the mystery of why the network throughput of a single process on some machines was very much slower than on others. Modern CPU have quite a few cores (our compute systems have 2 CPUs with 96 cores each). They also support a technology called hyperthreading where a single core can process two concurrent threads. This works well for some workloads although I have never been quite convinced by that on HPC loads. Anyway, looks like hyperthreading also doesn't play nice with 100GbE. I have to experiment some more but I suspect we will have to switch off hyperthreading. Something to play with tomorrow.

We got some very pretty clouds today and the view from the office was rather fine.