RE: [Xen-ia64-devel] [PATCH][RFC]discontig memory support

["Tian, Kevin" <kevin.tian@xxxxxxxxx>]

>From: Kouya SHIMURA [mailto:kouya@xxxxxxxxxxxxxx]
>Sent: 2006年4月4日 10:18
>Tian, Kevin writes:
> > If you can measure the TLB miss ratio of accessing vmemmap, that
>could
> > help a lot to make right choice.
>
>My guess is that the TLB miss ratio is quite worse because
>vmemmap(frame_table) is usually accessed only by grant table
>operation
>and the locality seems to be less.
>
>I think if the overhead of TLB miss would be enough light, the
>performance of VIRTUAL_MEM_MAP is better than SPARCEMEM. But
>SPARCMEM
>is more extensible to implement other functions.

One is grant table operation, another is the memory interface used by 
balloon driver, DMA, etc where such relationship will be queried. I'm not 
sure about frequency of grant table operations, but virtual drivers depend 
on that interface heavily, especially for vnif. When later those stuffs are 
ready to test, then you may measure whether the access to vmemmap is 
efficient.

A rough thought is, if most access to vmemmap results a TLB miss, and 
then go to physical mode, walk the tree page table with TLB insertion, 
back to virtual mode, it will become more efficient to walk the page table 
directly which would then look like SPARSEMEM. Or else vmemmap is 
the one as you said if overhead of TLB miss doesn't matter in the whole 
path. All depends on data. :-)

> > >[TODO]
> > >  * cleanup #ifdefs of CONFIG_VIRTUAL_MEM_MAP
>It means that CONFIG_VIRTUAL_MEM_MAP originated in Linux are still
>scattered in source code and I will do cleanup.

Good, and sorry for my misunderstanding.

Thanks,
Kevin

_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel