3 ms·

Yeah, we iterated a bunch on this question, and I still don't think we have it quite right. On the REPL side, we've had a few prototyped out [1] out over the y

• vegabook 10y ago
  Python (Numpy/Pandas really) is our go to for all the usual reasons on exploratory with some R (but R is becoming too slow for our growing data sets). What we're really missing is a language that does parallel computing, not necessarily only SIMD (Cuda style), but something that can also be more flexible on algorithms which have step dependencies. Scala/Spark works here, but it's more about big business batch data in our view, whereas we are about big non-linear optimizations in finance (fixed income curve fitting on huge numbers of bonds), and a further requirement for soft real time. We're very very excited about the Knight's Landing Xeon Phi, which will, we think, give us good performance on SIMD, and much more algorithmic flexibility because it can also be treated as a 288-core Xeon (hopefully). I think Rust could be a very good fit here because I predict that these "mega-core" chips will become very important. We can tolerate very little latency between processes so we basically cannot go multi-node too easily. We are also relatively small so don't have super-computer budgets.