RE: [Xen-ia64-devel] RE: rid virtualization

["Magenheimer, Dan (HP Labs Fort Collins)" <dan.magenheimer@xxxxxx>]

>       It is still too early for Xen/IA64 to do performance 
> mesurement now as there are so many stability issues so far.

Perhaps you are right.  I expect that it would take a big
database application to show any difference between the
two algorithms.

> As Matt has a lot of experience in LVHPT hash algorithm 
> (Linux LVHPT previous maintainer) and we agree his analysis 
> is correct, right? 

OK.  But then let's do it right by reversing all the bits
so that we don't have to change it yet again in the future.

See:
http://lists.xensource.com/archives/html/xen-ia64-devel/2005-09/msg00102.html

>       Then back to the coding style for mangling, I suggest 
> we should keep one place to do all the mangling like the C 
> code did now, but for all those FAST_*, shoud we change that 
> to a unique MACRO or a command function? 
> Eddie

I think there is only one place where mangling is done in
assembly (hyper_set_rr), but I could be wrong.  In any
case, I agree that the definitions should be in one place.

We could put an assembly macro in region_reg.h (using
ifdef __ASSEMBLER__ around the macro and ifndef __ASSEMBLER__
around all C code.

Dan
 
> -----Original Message-----
> From: Magenheimer, Dan (HP Labs Fort Collins) 
> [mailto:dan.magenheimer@xxxxxx] 
> Sent: 2005å11æ19æ 23:24
> To: Dong, Eddie; Matt Chapman
> Cc: xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
> Subject: RE: [Xen-ia64-devel] RE: rid virtualization
> 
> > > I think it would be worth changing the Xen mangling so that it
> > > switches bytes 1 and 2 instead of 1 and 3, and seeing if 
> that makes
> > > an improvement.
> > > 
> > > Matt
> > 
> > Dan:
> >     I noticed the latest code is still mangling byte 1 with  bytes 3
> > so far. Don't you want to switch to byte 1 and byte 2 
> mangling that is
> > obvious better than current one?
> > Thx,eddie
> 
> Do you have any benchmark results comparing the two (preferably
> with all the FAST_* options turned on in hyperprivop.S)?  Also,
> Matt observed that reversing all the bits should be even better.
> It would be good to benchmark that too.
> 
> If you are looking at this code, please also look to see if
> you find any places where rid mangling should be occuring
> but is not (e.g. metaphysical mode).  I think I saw a place
> where this might be a bug when I was working on fixing another
> bug, but don't recall where it was.
> 
> Thanks,
> Dan
> 

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