> One major limitation of malloc (and even the best implementations like jemalloc and dlmalloc) is that they try to use a single allocator for each data structure. This is a mistake: A huge performance gain can be had by using a separate allocator for each of your data structures — or rather, for each of your data usage patterns.
I stopped reading the article at this point. This statement has been disproven repeatedly over the years with Hoard and jemalloc. It is counter-intuitive but the data backs it up.
Custom per-data-structure allocators can fragment the global memory arena and cause more CPU cache misses as result of the extra code involved. The latest/greatest malloc/free implementations use a myriad of optimizations to achieve speed improvements that a custom allocator implementation would rarely use.
It's not an accident that jemalloc is so widely used in major applications - it works extremely well.