[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Fri, 2 Feb 2018 13:26:25 -0800
From: Frank Rowand <frowand.list@...il.com>
To: Chintan Pandya <cpandya@...eaurora.org>,
Rob Herring <robh+dt@...nel.org>
Cc: devicetree@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] of: cache phandle nodes to decrease cost of
of_find_node_by_phandle()
On 02/01/18 21:59, Chintan Pandya wrote:
>
>
> On 2/2/2018 12:40 AM, Frank Rowand wrote:
>> On 02/01/18 02:31, Chintan Pandya wrote:
>>>
>>>>> Anyways, will fix this locally and share test results.
>>>>
>>>> Thanks, I look forward to the results.
>>>>
>>>
>>> Set up for this time was slightly different. So, taken all the numbers again.
>>>
>>> Boot to shell time (in ms): Experiment 2
>>> [1] Base : 14.843805 14.784842 14.842338
>>> [2] 64 sized fixed cache : 14.189292 14.200003 14.266711
>>> [3] Dynamic freeable cache : 14.112412 14.064772 14.036052
>>>
>>> So, [3] (this patch) looks to be improving 750ms (on avg from base build).
>>>
>>
>> Is this with the many debug options enabled? If so, can you repeat with
>> a normal configuration?
>
> Could you share me the point of doing this experiment in perf mode ?
You had mentioned earlier in another thread:
My recent results were taken on debug_defconfig which has many performance
slowing code. So, gap between base-build and w/ the test patches would be
more than the actual production build.
If you measure a large performance gain with a debug configuration, that
may not represent the actual gain you will get with a production
configuration, as you noted.
My question was trying to determine whether the numbers reported above
are for a debug configuration or a production configuration. And if
not a production configuration, I was requesting the numbers for a
production configuration. If the production configuration does not
show a significant boot time reduction from the patch then there is
less justification for adding complexity to the existing code. I
prefer to use simpler data structures and algorithms __if__ extra
complexity does not provide any advantage. The balance between
complexity and benefits is a core software engineering issue.
> I don't have a set up for taking these numbers in perf mode. For
> that, I need to ask some other team and round trip follow ups. In my
> set up, I rely on serial console logging which gets disabled in perf
> mode.
>
>>
>> Thanks,
>>
>> Frank
>>
>
> Chintan
Powered by blists - more mailing lists