On 05/11/2015 04:53 PM, Pierre Thierry wrote:
Scribit salvador fandino dies 11/05/2015 hora 16:00:
Considering those results I would choose 200 as the best cut point because it is where SBCL and CCL perform best without incurring in a great penalty for ECL and GCL (~40%). Note also that performance degrades quite fast for SBCL and CCL once the optimal cut point is surpassed.
Why not use a different cut point for implementations where we know an optimal one?
IMO, those optimal values for the cut-point I have found in GCL and ECL (and unless some hidden issue is discovered in my code) are just revealing some bottleneck on that implementations. There is no reason, for the divide-and-conquer side to be so relatively slow.
Actually, I have found that compiling the function on GCL just brings the cut-point around the 200 mark too and on ECL to around 1000.