Hi,
I've been following the Transparent Hugepage patch set over the last few
weeks and have been experimenting with leveraging the functionality in
VMware's Linux virtualization software (eg: Workstation, Player). The
initial results have been extremely positive.
We have a standardized VM workload that we use to demonstrate the
difference that large page support makes - it's a Linux VM running
the SPECjbb benchmark in a JVM that has itself been configured to
use large pages through hugetlbfs - as such it's designed to give
pretty much best cases performance improvements.
In this benchmark, I've been seeing a 22% improvement in the operations
per second reported by SPECjbb.
We've done previous experiments with hugetlbfs but it really isn't
usable for backing VMs, so if we're ever going to be able to offer large
page support on Linux, it will require THP or something with comparable
functionality.
Particularly we appreciate:
* Not having to allocate large pages up front
* Being able to allocate small and large pages from the same source
* Being able to promote and demote pages as necessary
So, FWIW, THP really does help with virtualization workloads!
Thanks,
--phil