Xen devel - Mar 2010 - [PATCH] x86/hvm: accelerate IO intercept handling

Hi,

currently we go through the emulator every time a HVM guest does an I/O 
port access (in/out). This is unnecessary most of the times, as both VMX 
and SVM provide all the necessary information already in the VMCS/VMCB. 
String instructions are not covered by this shortcut, but they are quite 
rare and we would need to access the guest memory anyway.
This patch decodes the information from VMCB/VMCS and calls a simple 
handle_mmio wrapper. In handle_mmio() itself the emulation part will 
simply be skipped, this approach avoids code duplication.
Since the vendor specific part is quite trivial, I implemented both the 
VMX and SVM part, please check the VMX part for sanity.

I simply boot tested both versions and ran some simple benchmarks.
A micro benchmark (hammering an I/O port in a tight loop) shows a 
significant performance improvement (down to 66% of the time needed to 
handle the intercept on a K8, measured in the guest with TSC).
Even with reading a 1GB file from an emulated IDE harddisk (Dom0 cached) 
I could get a  4-5% improvement.
We found some guests (e.g. the TCP stack in some Windows version) which 
exercise the PM-Timer I/O port (0x1F48) very often (multiple 10,000 
times per second), these workloads also benefit from this patch.

Since this is not a regression, but only a performance improvement, I''d
suggest to apply to the post 4.0 tree.

Signed-off-by: Andre Przywara <andre.przywara@amd.com>

Regards,
Andre.

-- 
Andre Przywara
AMD-Operating System Research Center (OSRC), Dresden, Germany
Tel: +49 351 448 3567 12
----to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Andrew Bowd; Thomas M. McCoy; Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632


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

On 05/03/2010 14:36, "Andre Przywara" <andre.przywara@amd.com>
wrote:
> I simply boot tested both versions and ran some simple benchmarks.
> A micro benchmark (hammering an I/O port in a tight loop) shows a
> significant performance improvement (down to 66% of the time needed to
> handle the intercept on a K8, measured in the guest with TSC).
> Even with reading a 1GB file from an emulated IDE harddisk (Dom0 cached)
> I could get a  4-5% improvement.
> We found some guests (e.g. the TCP stack in some Windows version) which
> exercise the PM-Timer I/O port (0x1F48) very often (multiple 10,000
> times per second), these workloads also benefit from this patch.
By how much? I mean, the microbenchmark and 5% speedup on our
poor-man''s IO
path are not very interesting. Unless the speedup on the only
possibly-interesting workload you mention is significant, this whole
optimisation seems unnecessary.

 -- Keir



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

Keir Fraser wrote:
> On 05/03/2010 14:36, "Andre Przywara"
<andre.przywara@amd.com> wrote:
> 
>> I simply boot tested both versions and ran some simple benchmarks.
>> A micro benchmark (hammering an I/O port in a tight loop) shows a
>> significant performance improvement (down to 66% of the time needed to
>> handle the intercept on a K8, measured in the guest with TSC).
>> Even with reading a 1GB file from an emulated IDE harddisk (Dom0
cached)
>> I could get a  4-5% improvement.
>> We found some guests (e.g. the TCP stack in some Windows version) which
>> exercise the PM-Timer I/O port (0x1F48) very often (multiple 10,000
>> times per second), these workloads also benefit from this patch.
> 
> By how much? I mean, the microbenchmark and 5% speedup on our
poor-man''s IO
> path are not very interesting.
Educated estimation: Sysmark productivity should give about 0.5%, 
Passmark TCP localhost transfer on Windows 2008R2 should improve about 5%.

 > Unless the speedup on the only
> possibly-interesting workload you mention is significant, this whole
> optimisation seems unnecessary.
Actually it is missing enablement. What is the purpose of going through 
the emulator (mapping and walking guest page tables, reading guest 
instruction memory, decoding x86 code) when you don''t have to? KVM is 
implementing this for quite some time now.

And, after all, low hanging fruits are growing higher nowadays, so I''d 
consider an even modest performance improvement for _every_ machine by 
just a software patch a valuable thing.

Regards,
Andre.

-- 
Andre Przywara
AMD-Operating System Research Center (OSRC), Dresden, Germany
Tel: +49 351 448 3567 12
----to satisfy European Law for business letters:
Advanced Micro Devices GmbH
Karl-Hammerschmidt-Str. 34, 85609 Dornach b. Muenchen
Geschaeftsfuehrer: Andrew Bowd; Thomas M. McCoy; Giuliano Meroni
Sitz: Dornach, Gemeinde Aschheim, Landkreis Muenchen
Registergericht Muenchen, HRB Nr. 43632


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

On 05/03/2010 15:49, "Andre Przywara" <andre.przywara@amd.com>
wrote:
>> Unless the speedup on the only
>> possibly-interesting workload you mention is significant, this whole
>> optimisation seems unnecessary.
> Actually it is missing enablement.
Yes it''s very important to use *every* CPU feature. ;-)

Well, the patch looks okay, and it may speed up things we care about a
percent or two I guess. I will look at it properly after 4.0 and I''m
sure we
can get it in the tree.

 -- Keir



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