Re: [PATCH/RFC] Shared page tables

[Brian Twichell]

Hugh Dickins wrote:

> On Thu, 5 Jan 2006, Dave McCracken wrote:
>  
>
> > Here's a new version of my shared page tables patch.
> >
> > The primary purpose of sharing page tables is improved performance for
> > large applications that share big memory areas between multiple processes.
> > It eliminates the redundant page tables and significantly reduces the
> > number of minor page faults.  Tests show significant performance
> > improvement for large database applications, including those using large
> > pages.  There is no measurable performance degradation for small processes.
> >
> > This version of the patch uses Hugh's new locking mechanism, extending it
> > up the page table tree as far as necessary for proper concurrency control.
> >
> > The patch also includes the proper locking for following the vma chains.
> >
> > Hugh, I believe I have all the lock points nailed down.  I'd appreciate
> > your input on any I might have missed.
> >
> > The architectures supported are i386 and x86_64.  I'm working on 64 bit
> > ppc, but there are still some issues around proper segment handling that
> > need more testing.  This will be available in a separate patch once it's
> > solid.
> >
> > Dave McCracken
> >    
>
> The locking looks much better now, and I like the way i_mmap_lock seems
> to fall naturally into place where the pte lock doesn't work.  But still
> some raciness noted in comments on patch below.
>
> The main thing I dislike is the
> 16 files changed, 937 insertions(+), 69 deletions(-)
> (with just i386 and x86_64 included): it's adding more complexity than
> I can welcome, and too many unavoidable "if (shared) ... else ..."s.
> With significant further change needed, not just adding architectures.
>
> Worthwhile additional complexity?  I'm not the one to judge that.
> Brian has posted dramatic improvments (25%, 49%) for the non-huge OLTP,
> and yes, it's sickening the amount of memory we're wasting on pagetables
> in that particular kind of workload.  Less dramatic (3%, 4%) in the
> hugetlb case: and as yet (since last summer even) no profiles to tell
> where that improvement actually comes from.
>
>  
Hi,

We collected more granular performance data for the ppc64/hugepage case.

CPI decreased by 3% when shared pagetables were used.  Underlying this was a
7% decrease in the overall TLB miss rate.  The TLB miss rate for hugepages
decreased 39%.  TLB miss rates are calculated per instruction executed.

We didn't collect a profile per se, as we would expect a CPI improvement
of this nature to be spread over a significant number of functions,
mostly in user-space.

Cheers,
Brian


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/