On Thursday, June 11, 2015 at 2:53:36 PM UTC+2 Egon wrote: > 1. In which cases a cluster of say 4 (or 10 or 100 for instance) Raspberry >> Pi mini computers can be more cost-effective than a single computer with >> the same amount of cores (does the cost of communicating the data between >> the computers via the network not outweigh the fact that they car run tasks >> simultaneously) ? >> > > The general answer is Amdahl's Law ( > http://en.wikipedia.org/wiki/Amdahl%27s_law), of course it's not always > applicable ( > http://www.futurechips.org/thoughts-for-researchers/parallel-programming-gene-amdahl-said.html). > > When moving things to multiple-computers you'll get a larger overhead in > communication when compared to a single-computer, at the same time you may > reduce resource-contention for disk, RAM (or other resources). So depending > where your bottlenecks are, it could go either way... >
Yes, and also note that super-computers often use special network protocols/technologies which support so called "Remote direct memory access" (RDMA) [1], such as Infiniband [2], to get acceptable performance for high-performance multi-core computations across compute nodes. Infiniband cards are pretty expensive as far as I know, so will probably outweigh the benefits of buying a lot of RPis. I'd still be interested to hear if anybody knows about new developments on MPI for Go (for HPC use cases if nothing else)? :) [1] https://en.wikipedia.org/wiki/Remote_direct_memory_access [2] https://en.wikipedia.org/wiki/InfiniBand Best Samuel -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/d1d39602-e48e-4c2e-909b-a85d0d7e81ban%40googlegroups.com.
