RE: [PATCH] Optimizing x86-64 xenlinux using global pagesfor user mode (take 2)

RE: [PATCH] Optimizing x86-64 xenlinux using global pagesfor user mode (take 2)

[Ian Pratt]

> The previous patch had a problem with the builder, which constructs
the
> initial page tables with the USER bit on. And Xen couldn't distinguish
> kernel or use pages. It was partially checked avoiding the bug, and
it's
> finished with this patch.

Thanks. Have you any updated benchmark numbers as the bug could have
been giving quite a performance boost before.

Thanks,
Ian

 
>  linux-2.6-xen-sparse/include/asm-i386/mach-xen/asm/hypervisor.h |
4
> ++
>  linux-2.6-xen-sparse/include/asm-x86_64/mach-xen/asm/pgtable.h  |
19
> +++++-----
>  tools/libxc/xc_linux_build.c                                    |
3
> +
>  tools/libxc/xg_private.h                                        |
3
> +
>  xen/arch/x86/domain_build.c                                     |
2
> -
>  xen/arch/x86/flushtlb.c                                         |
7
> +++
>  xen/arch/x86/mm.c                                               |
12
> ++++--
>  xen/arch/x86/x86_64/traps.c                                     |
2
> -
>  xen/include/asm-x86/flushtlb.h                                  |
8
> +++-
>  xen/include/asm-x86/x86_64/page.h                               |
8
> +++-
>  10 files changed, 50 insertions(+), 18 deletions(-)
> 
> Signed-off-by: Jun Nakajima <jun.nakajima@intel.com>
> 
> Jun
> ---
> Intel Open Source Technology Center

Re: [PATCH] Optimizing x86-64 xenlinux using global pagesfor user mode (take 2)

[Andrew Theurer]

Ian Pratt wrote:
>> The previous patch had a problem with the builder, which constructs the
>> initial page tables with the USER bit on. And Xen couldn't distinguish
>> kernel or use pages. It was partially checked avoiding the bug, and it's
>> finished with this patch.
>
> Thanks. Have you any updated benchmark numbers as the bug could have
> been giving quite a performance boost before.

FWIW, I had been running more extensive testing on changeset 11225 
(before global-bit) and 11229 (global-bit with revert/fix for dom0 
builder), and they do not match the initial testing I did with rev 11118 
+ global_1 patch.  In fact, I am not seeing really any improvement.  I 
am not sure what is wrong.  Should I just test 
xen-unstable-tip+newest_patch and see where we stand?

-Andrew Theurer

RE: [PATCH] Optimizing x86-64 xenlinux using global pagesfor user mode (take 2)

[Nakajima, Jun]

Andrew Theurer wrote:
> Ian Pratt wrote:
>>> The previous patch had a problem with the builder, which constructs
>>> the initial page tables with the USER bit on. And Xen couldn't
>>> distinguish kernel or use pages. It was partially checked avoiding
>>> the bug, and it's finished with this patch.
>> 
>> Thanks. Have you any updated benchmark numbers as the bug could have
>> been giving quite a performance boost before.

Looks like part of the boost was realized by the bug, but I still see
consistent and visible improvements with lmbench. If I turn on the
global bits for the kernel pages as well, I see more improvements, which
are comparable with or better than before. I think we need to do macro
benchmarks to see how this helps. 

> 
> FWIW, I had been running more extensive testing on changeset 11225
> (before global-bit) and 11229 (global-bit with revert/fix for dom0
> builder), and they do not match the initial testing I did with rev
> 11118 + global_1 patch.  In fact, I am not seeing really any
> improvement.  I am not sure what is wrong.  Should I just test
> xen-unstable-tip+newest_patch and see where we stand?
> 

Yes, please. You shouldn't have observed any performance difference
between 11225 and 11229 because the changeset 11226 did not enable the
global bit. 

> -Andrew Theurer

Jun
---
Intel Open Source Technology Center

Re: [PATCH] Optimizing x86-64 xenlinux using global pagesfor user mode (take 2)

[Andrew Theurer]

Nakajima, Jun wrote:
> Andrew Theurer wrote:
>   
>> Ian Pratt wrote:
>>     
>>>> The previous patch had a problem with the builder, which constructs
>>>> the initial page tables with the USER bit on. And Xen couldn't
>>>> distinguish kernel or use pages. It was partially checked avoiding
>>>> the bug, and it's finished with this patch.
>>>>         
>>> Thanks. Have you any updated benchmark numbers as the bug could have
>>> been giving quite a performance boost before.
>>>       
>
> Looks like part of the boost was realized by the bug, but I still see
> consistent and visible improvements with lmbench. If I turn on the
> global bits for the kernel pages as well, I see more improvements, which
> are comparable with or better than before. I think we need to do macro
> benchmarks to see how this helps. 
>
>   
>> FWIW, I had been running more extensive testing on changeset 11225
>> (before global-bit) and 11229 (global-bit with revert/fix for dom0
>> builder), and they do not match the initial testing I did with rev
>> 11118 + global_1 patch.  In fact, I am not seeing really any
>> improvement.  I am not sure what is wrong.  Should I just test
>> xen-unstable-tip+newest_patch and see where we stand?
>>
>>     
>
> Yes, please. You shouldn't have observed any performance difference
> between 11225 and 11229 because the changeset 11226 did not enable the
> global bit. 

OK, then that would explain it :)  I'll run tip+ your latest patch.

-Andrew Theurer