[v1] Optimize load uCode performance

[edk2] [Patch 0/3] Optimize load uCode performance

Posted by Eric Dong 5 years, 9 months ago

Use below three rules to optimize load uCode performance:
1. Let BSP relocate uCode from flash to memory for better performance.
2. BSP caches the CPU ID and address of uCode so AP doesn\ufffdt need to look 
   for the uCode again if the CPU ID is same as BSP\ufffds.
3. Only apply uCode in one thread of a core when hyper threading is enabled.

Test:
Use an sample platform which has 1 socket, 4 core, 8 threads, the CpuMpPei
driver cost time reduce from 108.4ms to 27.2ms

Eric Dong (3):
  UefiCpuPkg/MpInitLib: Relocate uCode to memory to save time.
  UefiCpuPkg/MpInitLib: Use BSP uCode for APs if possible.
  UefiCpuPkg/MpInitLib: Load uCode once for one core.

 UefiCpuPkg/Library/MpInitLib/Microcode.c | 43 +++++++++++++++++++++++++++++---
 UefiCpuPkg/Library/MpInitLib/MpLib.c     | 17 ++++++++++---
 UefiCpuPkg/Library/MpInitLib/MpLib.h     | 11 ++++++--
 3 files changed, 63 insertions(+), 8 deletions(-)

-- 
2.15.0.windows.1

_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel

Re: [edk2] [Patch 0/3] Optimize load uCode performance

Posted by Laszlo Ersek 5 years, 9 months ago

Hi Eric,

On 07/11/18 13:07, Eric Dong wrote:
> Use below three rules to optimize load uCode performance:
> 1. Let BSP relocate uCode from flash to memory for better performance.
> 2. BSP caches the CPU ID and address of uCode so AP doesn’t need to look 
>    for the uCode again if the CPU ID is same as BSP’s.
> 3. Only apply uCode in one thread of a core when hyper threading is enabled.
> 
> Test:
> Use an sample platform which has 1 socket, 4 core, 8 threads, the CpuMpPei
> driver cost time reduce from 108.4ms to 27.2ms
> 
> Eric Dong (3):
>   UefiCpuPkg/MpInitLib: Relocate uCode to memory to save time.
>   UefiCpuPkg/MpInitLib: Use BSP uCode for APs if possible.
>   UefiCpuPkg/MpInitLib: Load uCode once for one core.
> 
>  UefiCpuPkg/Library/MpInitLib/Microcode.c | 43 +++++++++++++++++++++++++++++---
>  UefiCpuPkg/Library/MpInitLib/MpLib.c     | 17 ++++++++++---
>  UefiCpuPkg/Library/MpInitLib/MpLib.h     | 11 ++++++--
>  3 files changed, 63 insertions(+), 8 deletions(-)

I'm putting off the regression-testing of this series until we come to a
conclusion on pushing Ray's patch first (see my message at
<http://mid.mail-archive.com/26d8afe6-f89b-ddfc-0e8b-6acdc311f121@redhat.com>).

Thanks!
Laszlo
_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel

Re: [edk2] [Patch 0/3] Optimize load uCode performance

Posted by Laszlo Ersek 5 years, 9 months ago

On 07/11/18 13:07, Eric Dong wrote:
> Use below three rules to optimize load uCode performance:
> 1. Let BSP relocate uCode from flash to memory for better performance.
> 2. BSP caches the CPU ID and address of uCode so AP doesn’t need to look 
>    for the uCode again if the CPU ID is same as BSP’s.
> 3. Only apply uCode in one thread of a core when hyper threading is enabled.
> 
> Test:
> Use an sample platform which has 1 socket, 4 core, 8 threads, the CpuMpPei
> driver cost time reduce from 108.4ms to 27.2ms
> 
> Eric Dong (3):
>   UefiCpuPkg/MpInitLib: Relocate uCode to memory to save time.
>   UefiCpuPkg/MpInitLib: Use BSP uCode for APs if possible.
>   UefiCpuPkg/MpInitLib: Load uCode once for one core.
> 
>  UefiCpuPkg/Library/MpInitLib/Microcode.c | 43 +++++++++++++++++++++++++++++---
>  UefiCpuPkg/Library/MpInitLib/MpLib.c     | 17 ++++++++++---
>  UefiCpuPkg/Library/MpInitLib/MpLib.h     | 11 ++++++--
>  3 files changed, 63 insertions(+), 8 deletions(-)
> 
> 
> 
> _______________________________________________
> edk2-devel mailing list
> edk2-devel@lists.01.org
> https://lists.01.org/mailman/listinfo/edk2-devel
> 

I also tried to regression test this with OVMF. (I know I requested some
changes for patch #1, however those would only affect behavior in
practice when memory allocation fails, so it makes sense for me to test
this version of the series as well.)

Unfortunately, the patch set doesn't apply; it fails with the first
patch already. I tried on top of current master, commit 0a563f3fecfd
("BaseTool: Fixed the incorrect cache key.", 2018-07-12).

Can you please post a v2 that is also rebased on current master?

Thanks
Laszlo
_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel

Re: [edk2] [Patch 0/3] Optimize load uCode performance

Posted by Dong, Eric 5 years, 9 months ago

Hi Laszlo,

> -----Original Message-----
> From: Laszlo Ersek [mailto:lersek@redhat.com]
> Sent: Thursday, July 12, 2018 5:58 PM
> To: Dong, Eric <eric.dong@intel.com>; edk2-devel@lists.01.org
> Subject: Re: [edk2] [Patch 0/3] Optimize load uCode performance
> 
> On 07/11/18 13:07, Eric Dong wrote:
> > Use below three rules to optimize load uCode performance:
> > 1. Let BSP relocate uCode from flash to memory for better performance.
> > 2. BSP caches the CPU ID and address of uCode so AP doesn’t need to look
> >    for the uCode again if the CPU ID is same as BSP’s.
> > 3. Only apply uCode in one thread of a core when hyper threading is
> enabled.
> >
> > Test:
> > Use an sample platform which has 1 socket, 4 core, 8 threads, the
> > CpuMpPei driver cost time reduce from 108.4ms to 27.2ms
> >
> > Eric Dong (3):
> >   UefiCpuPkg/MpInitLib: Relocate uCode to memory to save time.
> >   UefiCpuPkg/MpInitLib: Use BSP uCode for APs if possible.
> >   UefiCpuPkg/MpInitLib: Load uCode once for one core.
> >
> >  UefiCpuPkg/Library/MpInitLib/Microcode.c | 43
> +++++++++++++++++++++++++++++---
> >  UefiCpuPkg/Library/MpInitLib/MpLib.c     | 17 ++++++++++---
> >  UefiCpuPkg/Library/MpInitLib/MpLib.h     | 11 ++++++--
> >  3 files changed, 63 insertions(+), 8 deletions(-)
> >
> >
> >
> > _______________________________________________
> > edk2-devel mailing list
> > edk2-devel@lists.01.org
> > https://lists.01.org/mailman/listinfo/edk2-devel
> >
> 
> I also tried to regression test this with OVMF. (I know I requested some
> changes for patch #1, however those would only affect behavior in practice
> when memory allocation fails, so it makes sense for me to test this version of
> the series as well.)
> 
> Unfortunately, the patch set doesn't apply; it fails with the first patch already.
> I tried on top of current master, commit 0a563f3fecfd
> ("BaseTool: Fixed the incorrect cache key.", 2018-07-12).
> 
> Can you please post a v2 that is also rebased on current master?
> 

I just send v2 patches base on the latest codebase, please verify with the newer one.

> Thanks
> Laszlo
_______________________________________________
edk2-devel mailing list
edk2-devel@lists.01.org
https://lists.01.org/mailman/listinfo/edk2-devel