Re: [boost] [Block Pointer] benchmark

Subject: Re: [boost] [Block Pointer] benchmark
From: Mathias Gaunard (mathias.gaunard_at_[hidden])
Date: 2011-05-25 13:06:38

On 25/05/2011 18:19, Phil Bouchard wrote:
> On 5/25/2011 5:10 AM, Mathias Gaunard wrote:
>
>> This is a fairly bad way to write a benchmark, and is likely to yield
>> somewhat incorrect results.
>
> It is based on a fixed amount of work and returns an estimate of time.
>
>> You should bench based on an external condition, such as time, rather
>> than making a fixed number of iterations.
>
> You mean: it should return the amount of iterations based on a fixed
> interval of time. I don't think it'll be as precise.

No, it should return the median time (or better yet, the median amount 
of cycles) taken by a single iteration, but you should do many 
iterations for a fixed interval of time.

This is necessary to prevent the compiler from reordering things between 
iterations or to unroll or optimize out the loop, since you have a 
non-predictable branch.