I think that 30% improvement you're referring to must be for just the i/o itself. Given that the i/o for the problem amounts to less than 1/1000th of the processing time, a 30% improvement is basically irrelevant. It doesn't help much to speed up something that's not even close to being a bottleneck. I think many of the best algorithm contest coders just use whatever i/o method uses the least code, keeps their answers simple and clean.