Re: O->A loan

[Matt Thomas <matt%3am-software.com@localhost>]

On Feb 19, 2012, at 7:52 PM, YAMAMOTO Takashi wrote:

> hi,
> 
>> yamt%mwd.biglobe.ne.jp@localhost (YAMAMOTO Takashi) wrote:
>>> hi,
>>> 
>>> i'd like to to either fix or remove O->A page loaning.
>>> 
>>> the biggest problem is the lack of users.  thus no way to test.
>>> any ideas?
>> 
>> After looking at O->A loaning code in yamt-pagecache branch, looking at
>> the benchmark results by Jaime (some data was sent privately, I think)
>> and contemplating about this mechanism - I tend to think that O->A loaning
>> should be removed.  It is a significant code complexity in UVM with some
>> special casing and benchmarks indicate that the benefit is somewhat small.
>> 
>> Does not seem to be worth, does it?
> 
> with this particular use alone, i tend to agree.
> 
> fwiw, the following is probably one of the best cases for loaned read.
> 
> YAMAMOTO Takashi
> 
> without loaned read:
> 
>       kuma% dd if=/netbsd bs=1m|wc -c
>       14+1 records in
>       14+1 records out
>       14768691 bytes transferred in 0.038 secs (388649763 bytes/sec)
>        14768691
> 
> with loaned read:
> 
>       kuma% dd if=/netbsd bs=1m|wc -c
>       14+1 records in
>       14+1 records out
>       14768691 bytes transferred in 0.023 secs (642117000 bytes/sec)
>        14768691

But that is a relatively rare use and the added to complexity to UVM doesn't 
seem worth it to me.