RE: [Xen-ia64-devel] [PATCH] [Resend]Enable hash vtlb

["Xu, Anthony" <anthony.xu@xxxxxxxxx>]

>From: Alex Williamson 
>Sent: 2006?4?10? 22:33
>On Mon, 2006-04-10 at 22:07 +0800, Xu, Anthony wrote:
>> Hi Alex,
>>
>> Below data is got based on changeset 8489.
>>
>> System:
>>      Tiger 4
>>      4G RAM (2GB available to xen)
>>      Montecito 1.4GHz dual core dual thread.
>>      DomU 512M RAM
>
>   Adding memory might be interesting.  Perhaps there's a range of
>memory where hash vtlb performs better.  For my dom0 testing, I booted
>w/ dom0_mem=768M (as this seemed to be the most I could do).  For domU
>testing, my xen config file specified 768MB of memory and dom0 was
>booted w/ the standard 512MB.  This still left a difference of ~40MB as
>reported by free once each domain is booted (domU has more memory).
>Swap was disabled in both cases and all extraneous daemons were stopped.
>
If we configure domU with memory 256MB, domU will complain "at least 256M 
is needed." 
Yes there should a best ratio of memory size of domU and size of VHPT.

>> bare metal (UP):
>> Total TimeBuild Time 2  Build Time 1
>> =====================================================================
>> 4008 Second  2004 Second  1995 Second
>> =====================================================================
>>
>> domU w/o hash vtlb
>> Total TimeBuild Time 2  Build Time 1
>> =====================================================================
>> 3966 Second  1976 Second  1978 Second
>> =====================================================================
>
>   I don't understand this result.  I was surprised to see domU perform
>better than dom0 in my testing, but I can't see how domU could perform
>better than bare metal.  Perhaps 512MB is insufficient for kernel
>builds.  You may be disproportionately benefiting from dom0's buffer
>cache.
>
I think there maybe two reasons.
1. As you said, domU benefits from dom0's buffer cache. There are somewhat 
parallel executions. DomU is response of compilation, Dom0 is response of 
read/write of disk.
2. The services running on Dom0 or DomU are less than that on native machine.


>> domU w/ hash vtlb
>> Total Time      Build Time1  Build Time2
>> =====================================================================
>> 3959 Second  1970 Second  1975 Second
>> =====================================================================
>>
>>
>> DomU can still get better performance after applying hash_vtlb patch.
>> This performance gain is based on 2% degradation of Dom0.(I don't get
>performance
>> data on Dom0 this time)
>>
>> The attachment is the script which I used to get kernel build performance.
>> Usage Example,
>> ./make_kernel.sh    2    /root/linux-2.6.16.tar.bz2
>>                   (times of build)   (absolute path)
>
>   The attachment seems to have been lost to a virus scanner.  My test
>was simply:
>
Almost same

Thanks
Anthony

># make clean
># time make > /dev/null 2>&1
>repeat
>
>Thanks,
>
>       Alex
>
>--
>Alex Williamson                             HP Linux & Open Source Lab

_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel