By: David Kanter (dkanter.delete@this.realworldtech.com), September 17, 2007 7:51 am
Room: Moderated Discussions
Jonathan Kang (johnbk@gmail.com) on 9/17/07 wrote:
---------------------------
>A very nice article. Thank you for taking the time to write this.
>
>One thing that drew my attention was remote prefetching. In the article, you covered
>the situation when a remote device would write to a processor cache directly. Even
>though this goes through the processor itself, doesn't this >introduce situations where cache thrashing could occur?
Absolutely - that's one of the reasons I would expect that a cache might refuse the prefetch.
>Suppose the processor switches context during a "DMA" >write. Is there a special
>cache tag that denotes that such regions shall not be >touched?
I'm not entirely sure how it works. That technique was straight from a patent.
>Also interesting would be how complicated prefetch >"patterns" could be.
I don't know the answer to this at all. I can try and find out.
DK
---------------------------
>A very nice article. Thank you for taking the time to write this.
>
>One thing that drew my attention was remote prefetching. In the article, you covered
>the situation when a remote device would write to a processor cache directly. Even
>though this goes through the processor itself, doesn't this >introduce situations where cache thrashing could occur?
Absolutely - that's one of the reasons I would expect that a cache might refuse the prefetch.
>Suppose the processor switches context during a "DMA" >write. Is there a special
>cache tag that denotes that such regions shall not be >touched?
I'm not entirely sure how it works. That technique was straight from a patent.
>Also interesting would be how complicated prefetch >"patterns" could be.
I don't know the answer to this at all. I can try and find out.
DK