Re: [Xen-ia64-devel] [PATCH 0/2][RFC] set rr0 to rr4 hyperprivop

[Isaku Yamahata <yamahata@xxxxxxxxxxxxx>]

On Tue, Jul 31, 2007 at 09:41:10AM -0600, Alex Williamson wrote:
> On Mon, 2007-07-30 at 14:59 +0900, Isaku Yamahata wrote:
> > This patch introduce set rr0 to rr4 hyperprivop to reduce context switch
> > overhead.
> > 
> > Linux/ia64 changes region registers from rr0 to rr4 when context switch
> > so that XenLinux issues 5 set_rr hyperprivops.
> > This patch introduces the hyperprivops to change and reduce the cost
> > to about 1/4 - 1/5. This rage is very reasonable because its overhead
> > comes from hypercall trap cost.
> > 
> > [PATCH 1/2][RFC] set rr0 to rr4 hyperprivop: Xen side
> > [PATCH 2/2][RFC] set rr0 to rr4 hyperprivop: Linux side
> 
> Hi Isaku,

Hi. Thank you for your comment and test.


>    I haven't been able to measure a performance difference from this,

I did micro benchmark and kernel compile benchmark.
My kernel compile benchmark showed small gain, but I did only few times.
Reviewing it now, the resultant time varies much. Probably I should
have repeaded kernel compile more, so the results would have converged
probably.
Micro benchmark by repeating the hyperprivops without accessing
region 0-4 showed the difference hypercall const.
Probably mov rr[] cost didn't appear there. I guess the existing 
set_rr hypercall cost is already small enough compared to real
mov rr[] cost.

I expected small gain with this patch because mov rr[] is already
paravirtualized and mov rr[] virtualization path is hand assembly 
optimized on both domU and domVTi. But I was wrong.


> but is there any reason not to do it?  It looks like a standard
> hypercall batching technique and the paravirtualization impact to the
> guest is quite low.  Thanks,

Only the maintenance cost without gain.
Although I don't think that this patch would cause a pain because
this patch can be reverted anytime.
If you think it's appropriate, please apply them.

-- 
yamahata

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