[erlang-questions] At what point am I "playing compiler"

Toby Thain <>
Fri May 22 22:29:34 CEST 2009


On 20-May-09, at 4:27 PM, John Haugeland wrote:

>> Ah, but the *more common* behaviour is to optimise unnecessarily,  
>> not to
>> reject optimisation where it's merited (how would that fly for  
>> very long?)
>
> I wonder why you believe this. ...
>
> Do you have any data to back up this claim, that premature
> optimization is more common than inadequate forward design?
>
>

Data? Only my noticing that programmers prefer the seductions and  
small victories of micro-optimisation than the difficult but more  
lucrative payoffs of design. But ymmv.

>
>
>
>> If it's needed, it's needed. If it's not, it's not. That is a rule  
>> of thumb
>> that is essentially Knuth's rephrased.
>
> We disagree on this point.  I would recommend a reading of the text;

Few would disagree that if it's not *needed* it should not be done,  
since optimisation usually brings other costs.

In any case, Joe's email is, imho, a good summary of the general idea.

...
>
>
>> Your "counterexample" was obviously a
>> case that was "needed" because it's built into your requirements:  
>> "I would
>> prefer this to take days, rather than months".
>
> And that has essentially nothing to do with my commentary that
> platitudes don't scale mathematically.  It's also wrong: I made the
> adjustments after the first version of the system was entirely
> complete, on observing that my initial strategy was under-performing.
> ... what actually
> happened was the software was made, and then I had to go back and
> change it, because I did not perform the work that people want to
> refer to as "premature optimization".  That's weird because by the
> rules you're citing, I should be seen as having done it right;

Yes.

> I just
> built something that worked, and when underperformance was observed, I
> then went back and altered the system.  It was presented as an example
> contrary to the assumption-driven ratios of man hours and valuability
> and run repeat count.
>
> Unfortunately, despite that, you're trying to create a situation where
> I ignored a requirement that I had to go back and fix, which
> undermines the idea that these pieces of work shouldn't be done before
> completion.
>
> It's a simple yes or no: should I have done the refinement before the
> first version?

There's no black and white answer. I didn't mean to imply there was.

>
>  * If yes, where's the line between that and premature optimization?
> Does it involve knowing beforehand how everything will perform?  How
> do we pick up the talent to know these things?
>  * If no, then why isn't this a clear counterexample to prior claims
> of "only if it's run a million times" and the value of programmer
> hours for one-use code?
>
> From where I stand, it appears that you're trying to argue both sides
> of when to work on efficiency, in re: before and after performance
> characteristics are known.  I must misunderstand some part of your
> argument.  Perhaps you'd be so kind as to clarify?
>
>
>
>
>
>> Or, you never realised it could be made orders of magnitude faster  
>> until you
>> fixed a design flaw. (Optimisation by accident.)
>
> I'd appreciate it if you wouldn't argue with me on basis of guesses
> you're making about code you haven't seen.  Neither of your cases were
> correct: it wasn't built into my requirements: I didn't realize there
> would be a problem until I saw the lack of progress in the table, and
> the thing I fixed was the throughput overhead of having work done in
> the client application instead of in the database itself using stored
> procedures.
>
> I hope you won't amend the platitude again to continue to cope with
> the changing sands of finding out more information.  That's generally
> understood to be a red flag that the rule of thumb has failed.
>
>

Exactly, it's a *rule of thumb*, your heavy artillery in this  
argument is wasted.

>
>
>
>>> Skepticism in all things is the basis of deep understanding.
>>
>> ...and that includes skepticism about the reflex and attendant
>> "rationalisations" for micro-optimising early. (Which is just  
>> wasting a
>> different and more valuable resource.)
>
> Luckily, since I made no rationalization for micro-optimizing early,
> this isn't germane.  All I did was to point out that platitudes don't
> suit multiplicative extension by Moore's law.
>
> I believe that this belief that any planning regarding performance is
> early micro-optimization, and the groupthink that inevitably comes
> with it, is nearly as large a disservice to the competant programmer
> as actual premature optimization is.

Then I disagree, since imho there is nothing at all wrong with the  
simple process of "get it working; get it working correctly; get it  
fast [enough]."


--T



More information about the erlang-questions mailing list