Mike Cheponis wrote:
> I don't believe this is true.
> 
> If one were willing to measure transfer rates for a variety of sectors, 
> I suspect an exact map of the disk could be determined.
> 
> If the disk can tell you when it is re-mapping a sector, one could 
> adjust the algorithm.
> 
> I believe this uncertainty can be completely removed.

Is it worth having all this extra cruft to measure sector transfers in 
the file system for such a small payoff?  When does all this extra 
complexity become simply complexity and overhead for certain transfers? 
  The paper quoted by another poster says that performance can be 
increased by up to 50% for mid-sized requests (100 - 500KB).  That would 
only be useful if, like you say, we can get tight VM/FS integration 
where lots of transfers can be bundled up together.  And then, the disk 
extents would need to be contiguous.  Is it really worth it?  And what 
is *up to* 50%?  How many milliseconds, at best?

I don't believe it's worth it.  But, I am by no means an expert on 
disks, so if anyone knows anything I don't know... ;-)

> Yes.  Simulate it.

OK, thanks.  I'll look into it.  Another poster has pointed out 
something interesting that I'll check out.

> p.s. I have for years now advocated a tighter coupling between the VM 
> and a huge FS.  Others with less vision have tried to shoot me down.  
> Well, I'm still here, and we still need tight VM/FS integration.

I also think tight VM/FS integration would be good.  The NetBSD UVM is a 
brilliant piece of software engineering and it would be a shame to not 
use it.  Making good use of cache is obviously good for reducing 
unnecessary disk I/O.

--
James