> SuperMalloc exploits the fact that although physical memory is always precious, virtual address space on a 64-bit machine is relatively cheap. It allocates 2 MiB chunks which contain objects all the same size.
Hmm, this must not interact well with huge pages: either all those (1 huge page-sized) chunks actually take up physical memory or you're stuck with 4 KiB pages.
Transparent huge pages on Linux are a noticeable performance difference in practice. In one of my applications, a coworker discovered that dTLB misses were a significant source of slowness. He pointed out that we had a lot of "holes" in our memory allocation caused by the malloc implementation calling madvise on 4 KiB pages when they were empty, thus preventing them from being coalesced into a 2 MiB huge page (or causing them to be migrate back to small pages). Setting a commandline flag to prevent this decreased CPU usage by 7%.
Hmm, this must not interact well with huge pages: either all those (1 huge page-sized) chunks actually take up physical memory or you're stuck with 4 KiB pages.
Transparent huge pages on Linux are a noticeable performance difference in practice. In one of my applications, a coworker discovered that dTLB misses were a significant source of slowness. He pointed out that we had a lot of "holes" in our memory allocation caused by the malloc implementation calling madvise on 4 KiB pages when they were empty, thus preventing them from being coalesced into a 2 MiB huge page (or causing them to be migrate back to small pages). Setting a commandline flag to prevent this decreased CPU usage by 7%.