Messages in this thread | | | Date | Thu, 19 Sep 2013 12:17:27 -0500 | From | Alex Thorlton <> | Subject | Re: [PATCHv2 0/9] split page table lock for PMD tables |
| |
On Mon, Sep 16, 2013 at 02:25:31PM +0300, Kirill A. Shutemov wrote: > Alex Thorlton noticed that some massivly threaded workloads work poorly, > if THP enabled. This patchset fixes this by introducing split page table > lock for PMD tables. hugetlbfs is not covered yet. > > This patchset is based on work by Naoya Horiguchi. > > Changes: > v2: > - reuse CONFIG_SPLIT_PTLOCK_CPUS for PMD split lock; > - s/huge_pmd_lock/pmd_lock/g; > - assume pgtable_pmd_page_ctor() can fail; > - fix format line in task_mem() for VmPTE; > > Benchmark (from Alex): ftp://shell.sgi.com/collect/appsx_test/pthread_test.tar.gz > Run on 4 socket Westmere with 128 GiB of RAM.
Kirill,
I'm hitting some performance issues with these patches on our larger machines (>=128 cores/256 threads). I've managed to livelock larger systems with one of our tests (I'll publish this one soon), and I'm actually seeing a performance hit on some of the smaller ones. I'm currently collecting some results to show the problems I'm hitting, and trying to research what's causing the livelock. For now I just wanted to let you know that I'm seeing some issues. I'll be in touch with more details.
Sorry for the delayed response, I wanted to be really sure that I was seeing problems before I put up a red flag.
- Alex
| |