Re: [PATCH] VM: Implements the swap-out page-clustering technique

[Li Yu <raise.sail@gmail.com>]

Li Yu 写道:
> hamidreza jahanjou 写道:
>> 2008/9/8, Li Yu <raise.sail@gmail.com>:
>>   
>>> And the linear searching for whole address space of a vma is not good
>>> idea too, really.
>>>
>>>     
>>>> +	    if (!page_allowed_in_cluster(cursor_page, ci))
>>>> +		continue;
>>>>
>>>>       
>>
>> Thank you for the review and comments. You are right in that searching
>> the whole process address space is not a good idea. The idea is to
>> make the scan range flexible, thus normally, assuming that the code
>> has been well-tuned, only a very limited number of VMA's are scanned.
>>
>> In general, i think that the cluster size and the scan range should be
>> tuned depending on the backing storage characteristics and the
>> low-on-memory severity. Like any VM code, this one needs tuning.
>>
>> BTW, the code could be implemented much more elegantly if the VMA's
>> were connected using a doubly-linked list. I wonder if there are
>> enought other codes with the same situation to justify making it
>> doubly-linked.
>>   
>
> You are welcome!
>
> I think that scanning one vma each task is enough.
>
> The goal of swapoutclustering is to reduce disk heads movement while 
> swapping in, it is based on space/time locality. I think that 
> different vmas implied different using pattern, which also means they 
> have different in  program locality. Therefore, I think scanning multi 
> vmas per task is not a good idea.
>
> Maybe, we need a page cache searching like for anonymous mapping 
> instead of *linear*  searching in each vma.
>
> I also have an uncompleted implementation for this feature ^_^
>
> Good luck.
>