lkml.org 
[lkml]   [2010]   [Nov]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] vmscan: move referenced VM_EXEC pages to active list
From
Date
On Mon, 2010-11-01 at 21:41 +0800, Minchan Kim wrote:
> On Sun, Oct 31, 2010 at 11:58:52AM +0800, Shaohua Li wrote:
> > On Mon, 2010-10-25 at 06:52 +0800, Minchan Kim wrote:
> > > On Sat, Oct 23, 2010 at 7:51 AM, Mandeep Singh Baines <msb@chromium.org> wrote:
> > > > In commit 64574746, "vmscan: detect mapped file pages used only once",
> > > > Johannes Weiner, added logic to page_check_reference to cycle again
> > > > used once pages.
> > > >
> > > > In commit 8cab4754, "vmscan: make mapped executable pages the first
> > > > class citizen", Wu Fengguang, added logic to shrink_active_list which
> > > > protects file-backed VM_EXEC pages by keeping them in the active_list if
> > > > they are referenced.
> > > >
> > > > This patch adds logic to move such pages from the inactive list to the
> > > > active list immediately if they have been referenced. If a VM_EXEC page
> > > > is seen as referenced during an inactive list scan, that reference must
> > > > have occurred after the page was put on the inactive list. There is no
> > > > need to wait for the page to be referenced again.
> > > >
> > > > Change-Id: I17c312e916377e93e5a92c52518b6c829f9ab30b
> > > > Signed-off-by: Mandeep Singh Baines <msb@chromium.org>
> > >
> > > It seems to be similar to http://www.spinics.net/lists/linux-mm/msg09617.html.
> > > I don't know what it is going. Shaohua?
> > I should have sent the test result earlier but was offlined last week.
> > Here is my test result:
> > kernel1: base kernel + revert commit 8cab4754
> > kernel2: base kernel
> > kernel3: base kernel + my patch (similar like Mandeep's)
> > I'm using Fengguang's test of commit 8cab4754. But the test result isn't
> > stable, sometimes one kernel above has more majfault, but sometimes the
> > kernel has less majfault. This is true for all the above kernels.
> > Apparently kernel behavior changes (guess because of commit 64574746),
> > and vm_exec protect (even the vm_exec protect in active list) is not
> > important now with new kernel in Fengguang's test suite.
>
> Tend to agree.
> When I saw 64574746, I doubted 8cab4754's effectiveness.
> When we reviewed 8cab4754, there were many discussion.
> The thing I kept my mind was a trick of VM_EXEC.
> Someone can whip LRU by VM_EXEC hack intentionally.
> Apparently, It's bad.
>
> >
> > But on the other hand, if I add a new task into Fengguang's test suite.
> > The task produces a lot of used one file page read (sequential read a
> > large sparse file). Kernel2 has less majfault than kernel1, and kernel3
> > has even less majfault than kernel2, so kernel3 has best performance.
> > Basically the majfault number from kernel1 is 3x, kernel2 2x, kernel3
> > 1x. One issue is I'm afraid this isn't a typical desktop usage any more
> > (because of sequential read sparse file), so not sure if we can use this
> > test as a judgment to merge the patch.
>
> We can't make sure desktop doesn't has such workload and server also can have
> such workload. I mean if it enhance VM by general POV, we can merge it enoughly.
> In your testcase, Removing VM_EXEC test(ie, kernel 2) doesn't have biased.
> It means it's not the best but not worst, either.
> Although we can't get the best, we can remove VM_EXEC hack. It's not a bad deal.
> So how about removing VM_EXEC hack in this chance?
>
> I hope we revert VM_EXEC hack in this chance.
> Of course, before we discuss it, we can need more and detail data.
> I hope you could help for the number.
I'm thinking if we should revert VM_EXEC hack too. The headache is we
need find a typical workload which can convince people and get number
under the workload. Fengguang's test case hasn't enough non vm_exec file
pages I think.

Thanks,
Shaohua



\
 
 \ /
  Last update: 2010-11-02 01:33    [W:1.110 / U:0.000 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site