WF2 in POSIX and C++

[Christoph Bartoschek]

Hi,

I've written a C++ version that runs faster than the current ones.
With 8 worker threads it runs in about 5 minutes and 50 seconds. The
fastest run with 16 threads has:

real 5m5.724s
user 45m38.926s
sys 3m14.382s

A different implementation that is on average slower got the following
result:

real 4m40.956s
user 42m18.794s
sys 3m06.856s


I've written about my conclusions on the page
http://www.pontohonk.de/wide-finder2/wide-finder2.html
It would be nice, If someone could update the results page.

Greetings
Christoph Bartoschek

[Tim Bray]

On Sun, Jul 27, 2008 at 6:28 AM, Christoph Bartoschek
<goo...@pontohonk.de> wrote:
>
> I've written about my conclusions on the page
> http://www.pontohonk.de/wide-finder2/wide-finder2.html
> It would be nice, If someone could update the results page.

Go ahead. -Tim

[Alastair Rankine]

Christoph Bartoschek wrote:
> I've written about my conclusions on the page
> http://www.pontohonk.de/wide-finder2/wide-finder2.html
> It would be nice, If someone could update the results page.
>

Nice work Christoph, and certainly consistent with my own experiments.

I'm very curious as to why the POSIX I/O calls should result in such a
dramatic speedup.

To me, this raises questions about the effectiveness of the WF2
benchmark as an exercise in writing "computer programs to take advantage
of modern slow-clock-rate/many-core computers". To me it's more of an
exercise in picking the best (for the dataset and system) I/O model.
This has the potential (possibly already realised) to skew the results
towards implementations and runtimes that support this model -
irrespective of other benefits.

[Alex Morega]

On Jul 28, 2008, at 02:31 , Alastair Rankine wrote:

>
> Christoph Bartoschek wrote:
>> I've written about my conclusions on the page
>> http://www.pontohonk.de/wide-finder2/wide-finder2.html
>> It would be nice, If someone could update the results page.
>>
> Nice work Christoph, and certainly consistent with my own experiments.

I'll second that, it's nice to have an as-fast-as-possible reference
implementation.

> I'm very curious as to why the POSIX I/O calls should result in such a
> dramatic speedup.
>
> To me, this raises questions about the effectiveness of the WF2
> benchmark as an exercise in writing "computer programs to take
> advantage
> of modern slow-clock-rate/many-core computers". To me it's more of an
> exercise in picking the best (for the dataset and system) I/O model.
> This has the potential (possibly already realised) to skew the results
> towards implementations and runtimes that support this model -
> irrespective of other benefits.

That happens if you only look at the numbers. But the implementations
are open to (subjective) interpretation: the degree of readability/
maintainability, the amount of knowledge required to write/understand
them, etc.

-- Alex