Given the cost of TLB shootdowns, the API's should really be asynchronous. You don't always need all cores to see the same view of the page tables at the same time: https://www.usenix.org/legacy/events/vee05/full_papers/p46-c....

Efficient bulk TLB manipulation is also part of the secret sauce behind Azul's pauseless GC. AIUI they have a custom kernel module to provide more efficient virtual memory operations than the linux kernel APIs can provide.

What about Linux TLB Big Page [1]?

[1]: https://www.kernel.org/doc/Documentation/vm/hugetlbpage.txt

huge pages are useful to reduce TLB overhead too, but it's not the same as batched modifications.

TL;DR it pays off to invalidate a range of pages per inter-processor interrupt, rather than one page per interrupt, and dtrace can get the numbers to prove it.

What is TLB shootdown? http://stackoverflow.com/q/3748384/290085