This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH 1/2] benchtests: Memory walking benchmark for memcpy
- From: Carlos O'Donell <carlos at redhat dot com>
- To: siddhesh at sourceware dot org, libc-alpha at sourceware dot org
- Date: Wed, 4 Oct 2017 15:19:18 -0700
- Subject: Re: [PATCH 1/2] benchtests: Memory walking benchmark for memcpy
- Authentication-results: sourceware.org; auth=none
- References: <1505756414-12857-1-git-send-email-siddhesh@sourceware.org> <be10b3b8-6440-cacb-62e1-6e44559e7fca@redhat.com> <7d713462-4db7-bdb8-c42c-61da43ccbf9f@sourceware.org> <ea9cf4c3-e24f-ceae-2034-9a86368c0345@sourceware.org>
On 10/03/2017 11:53 PM, Siddhesh Poyarekar wrote:
> On Friday 22 September 2017 05:29 AM, Siddhesh Poyarekar wrote:
>> On Thursday 21 September 2017 11:59 PM, Carlos O'Donell wrote:
>>> I like the idea, and the point that the other benchmark eventually degrades
>>> into measuring L1 performance an interesting insight.
>>>
>>> I do not like that it produces total data rate not time taken per execution.
>>> Why the change? If time taken per execution was OK before, why not here?
>>
>> That is because it seems more natural to express string function
>> performance by the rate at which it processes data than the time it
>> takes to execute. It also makes comparison across sizes a bit
>> interesting, i.e. the data rate for processing 1MB 32 bytes at a time vs
>> 128 bytes at a time.
>>
>> The fact that "twice as fast" sounds better than "takes half the time"
>> is an added bonus :)
>
> Carlos, do you think this is a reasonable enough explanation? I'll fix
> up the output in a subsequent patch so that it has a 'throughput'
> property that the post-processing scripts can read without needing the
> additional argument in 2/2.
As the subsystem maintainer I defer to your choice here. I don't have a
strong opinion, other than a desire for conformity of measurements to
avoid confusion. If I could say anything, consider the consumer and make
sure the data is tagged such that a consumer can determine if it is time
or throughput.
--
Cheers,
Carlos.