Hybrid THP mechanism. Selective use of huge pages by hot applications

Currently, THP policies are used globally, for the entire system. This leads to memory fragmentation and memory waste. Main memory has increased a lot faster than TLB entries, and it will continue to do so. Given the limited TLB cache entries, this becomes a serious bottleneck for real world applications. Huge pages are supposed to resolve this issue, since the a single entry in the TLB can map a big chunk of memory. However, this eventually leads to memory fragmentation and eventually the system runs out of usable memory. Hence, most sysadmins and user space application suggest to disable huge pages

We introduce a hybrid page mechanism where hot applications can use huge pages transparently, while avoiding the entire system to use huge pages.

During the talk, we will show how we managed to decrease the huge page consumption as well as benchmarks on real applications.