In what sort of workload was a custom allocator the biggest performance win?