Re: [Xen-ia64-devel] kernel build performance data with dom0_VP modeopen - Isaku Yamahata - com.xensource.lists.xen-ia64-devel

7 messages in com.xensource.lists.xen-ia64-develRe: [Xen-ia64-devel] kernel build per...

From	Sent On	Attachments
Zhang, Jingke	22 May 2006 00:48
Magenheimer, Dan (HP Labs Fort Collins)	24 May 2006 10:22
Isaku Yamahata	24 May 2006 18:50
Magenheimer, Dan (HP Labs Fort Collins)	25 May 2006 10:28
Xu, Anthony	25 May 2006 18:30
Isaku Yamahata	25 May 2006 19:05
Isaku Yamahata	25 May 2006 20:20

Subject:	Re: [Xen-ia64-devel] kernel build performance data with dom0_VP modeopen
From:	Isaku Yamahata (yama...@valinux.co.jp)
Date:	05/24/2006 06:50:49 PM
List:	com.xensource.lists.xen-ia64-devel

On Wed, May 24, 2006 at 10:23:10AM -0700, Magenheimer, Dan (HP Labs Fort
Collins) wrote:

At first, I suspected that this degradation was entirely due to the fact that large (64MB?) kernel pages were now causing many more TLB/VHPT misses because they are fragmented into 16KB pages. However, I am seeing that domU performance is also reduced by ~3%, and domU had this fragmentation prior to VP.

Current the dom0vp model implementation flushes all tlb/vhpt when granted page is unmapped from dom0. I think this affects significantly on domU performance.

Jingke, have you performed the same measurement on domU? If so, can you post your results?

Anybody have any other ideas as to what might be causing this (major) performance degradation?

For dom0vp model, VHPT size is reduced to 64KBytes to avoid triggering software lock detection. For P=M model, its size remain 64MB. It would be interesting to measure on P=M model with 64KBtyes VHPT size. (Edit VHPT_SIZE_LOG2 in xen/include/asm-ia64/vhpt.h)

Thanks.

________________________________

From: xen-...@lists.xensource.com [mailto:xen-...@lists.xensource.com] On Behalf Of Zhang, Jingke Sent: Monday, May 22, 2006 1:49 AM To: xen-...@lists.xensource.com Subject: [Xen-ia64-devel] kernel build performance data with dom0_VP modeopen

Hi all,

I have done kernel build with dom0_vp mode open to compare with it close. The specific data as following:

Environment :

Memory:768M

Platform : tiger4

Cpu : Montecito C1 stepping

Compile cmdline : time make >/dev/null 2>&1

Configuration file : @arch/ia64/config/tiger_defconfig

time Real user sys

Cset 10027 (with dom0_vp open): 14m48.377s 13m43.960s 0m51.670s

Cset 10022 (with dom0_vp closed): 14m5.752s 13m13.200s 0m39.260s

Native : 13m55.572s 13m4.366s 0m35.205s

Performance :

Cset 10022: dom0/native =98.7%

Cset 10027:dom0/native=94.1%

From the result, we can see when dom0_vp mode open, the performance of KB has 4.6% degradation.

Thanks

Zhang Jingke