« Return to Thread: [vwnc] [vw76] Point>>hash still not good enough

Re: [vwnc] [vw76] Point>>hash still not good enough

by Reinout Heeck :: Rate this Message:

Reply to Author | View in Thread


On May 9, 2008, at 9:33 PM, Andres Valloud wrote:

> Reinout Heeck wrote:
>> Valloud, Andres wrote:
>>
>>> Hmmm... I think what is going on is that clustering is not letting  
>>> the
>>> dictionary do its job quickly.
>>>
>> Yes, it surprised me that this case was not fixed in the hashing  
>> overhaul.
>>
>>
>>
>
> For all hash functions there will be datasets that cause them to  
> behave
> badly.  This cannot be avoided.
>
>> I had hoped these [value judgment] days would be over...

Heh :-)

What I meant there was the days of worrying about clustered values.



More generally, before the hashing overhaul I had run into three  
problems repeatedly:
1) Poor performance of hashed collections holding strings.
2) Poor performance of hashed collections holding clustered integer  
values.
3) Poor abstraction of hash combination 'for dummies' leading to poor  
#hash implementations on domain objects.

An earlier conversation we had extinguished my hopes for any solution  
to 3) in the overhaul but I had fully expected both 1) and 2) to be  
resolved by the redesign. Now I get the impression only 1) has been  
addressed adequately.

You seem to be dismissing 2) as being a special case that needs not  
perform with the base collection implementations. Since this is a  
value judgment we can keep debating this, instead I can provide you a  
simple datapoint: my shop would be relieved when 2) is addressed.


I had a quick look (for the first time) at the hash analysis tool in  
order to measure performance of some hashed collection access  
patterns. I only found quality assessments of #hash implementations  
for integer based values but no benchmark tool to measure run time of  
collection accesses - it seems this is not supported. Is there some  
other tool I need to load in order to measure the performance of  
hashed collections?



Thanks for your suggestions regarding Point>>hash, I'll experiment  
with them when I'm back at work (bank holiday here today).

Cheers,

Reinout
-------

_______________________________________________
vwnc mailing list
vwnc@...
http://lists.cs.uiuc.edu/mailman/listinfo/vwnc

 « Return to Thread: [vwnc] [vw76] Point>>hash still not good enough

LightInTheBox - Buy quality products at wholesale price