Xen devel - Oct 2006 - SPEC cpu2000 will always fail for ia32 smp hvm guest

Hi all,
I find that SPEC cpu2k benchmark will always fail (9 of 10 times) for
ia32 smp hvm guest with 4 vcpus. I used 4 users parameter to run cpu2k
and assign 1G memory for the guest, but it will always fail when running
164.zip case and report warnings as the attachments. However, cpu2k
could be run successfully in native linux and xenU guest with the same
configuration. Anyone has ideas about this? Thank you.

Best regards,
Fan




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

Hi all,
I find that SPEC cpu2k benchmark will always fail (9 of 10 times) for
ia32 smp hvm guest with 4 vcpus. I used 4 users parameter to run cpu2k
and assign 1G memory for the guest, but it will always fail when running
164.zip case and report warnings as "out of memory". However, cpu2k
could be run successfully in native linux and xenU guest with the same
configuration. Anyone has ideas about this? Thank you.

Best regards,
Fan

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

BTW, if I assign 1.5G memory for hvm guest, cpu2k could pass with 4 users, but
at least 1.1G is not enough.

Best regards,
Fan

-----Original Message-----
From: xen-devel-bounces@lists.xensource.com
[mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Zhao, Fan
Sent: 2006年10月10日 10:29
To: xen-devel@lists.xensource.com
Subject: [Xen-devel] SPEC cpu2000 will always fail for ia32 smp hvm guest

Hi all,
I find that SPEC cpu2k benchmark will always fail (9 of 10 times) for
ia32 smp hvm guest with 4 vcpus. I used 4 users parameter to run cpu2k
and assign 1G memory for the guest, but it will always fail when running
164.zip case and report warnings as the attachments. However, cpu2k
could be run successfully in native linux and xenU guest with the same
configuration. Anyone has ideas about this? Thank you.

Best regards,
Fan

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

On 10/10/06 10:28, "Zhao, Fan" <fan.zhao@intel.com> wrote:
> BTW, if I assign 1.5G memory for hvm guest, cpu2k could pass with 4 users,
but
> at least 1.1G is not enough.
> 
> Best regards,
> Fan
When you tested with 1GB on native, did the native guest have a swap
partition? When testing on HVM, are you giving the HVM guest a swap
partition?

This does simply look like a straightforward out-of-memory condition.
Unsurprising if you''re doing big userspace tests. What *is* interesting
is
if it turns out that memory requirements running in an HVM guest are very
different from running on native. That shouldn''t be the case.

 -- Keir



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

Keir,
Thanks for your comments. Actually, I do not give a swap partition for both
native and hvm guest when I do the test.
One point which deserves to mention is that the tests for hvm guest and xenu
guest are based on the same image file, which helps to ensure that hvm and xenu
guests share the same testing environment. At this point, xenu guest
doesn''t have a swap partition either, but cpu2k could run successfully
with 1G memory in it.
Another point for reference is that cpu2k with 4 users could pass on ia32e hvm
guest with 1G memory, which also does not have a swap partition.

Best regards,
Fan

-----Original Message-----
From: Keir Fraser [mailto:Keir.Fraser@cl.cam.ac.uk] 
Sent: 2006年10月10日 18:00
To: Zhao, Fan; xen-devel@lists.xensource.com
Subject: Re: [Xen-devel] SPEC cpu2000 will always fail for ia32 smp hvm guest

On 10/10/06 10:28, "Zhao, Fan" <fan.zhao@intel.com> wrote:
> BTW, if I assign 1.5G memory for hvm guest, cpu2k could pass with 4 users,
but
> at least 1.1G is not enough.
> 
> Best regards,
> Fan
When you tested with 1GB on native, did the native guest have a swap
partition? When testing on HVM, are you giving the HVM guest a swap
partition?

This does simply look like a straightforward out-of-memory condition.
Unsurprising if you''re doing big userspace tests. What *is* interesting
is
if it turns out that memory requirements running in an HVM guest are very
different from running on native. That shouldn''t be the case.

 -- Keir

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

On 10/10/06 15:02, "Zhao, Fan" <fan.zhao@intel.com> wrote:
> Thanks for your comments. Actually, I do not give a swap partition for both
> native and hvm guest when I do the test.
> One point which deserves to mention is that the tests for hvm guest and
xenu
> guest are based on the same image file, which helps to ensure that hvm and
> xenu guests share the same testing environment. At this point, xenu guest
> doesn''t have a swap partition either, but cpu2k could run
successfully with 1G
> memory in it.
> Another point for reference is that cpu2k with 4 users could pass on ia32e
hvm
> guest with 1G memory, which also does not have a swap partition.
If you run ''cat /proc/meminfo'' on native and on HVM, do the
MemTotal figures
match up? If so, this must be an issue within the guest -- it would indicate
that we are providing the guest with the same amount of memory as was
available to it when running on native.

 -- Keir



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

Keir,
If I assign 1G memory for native, ''cat /proc/meminfo'' shows
the MemTotal is 1034672 KB, while ''cat /proc/meminfo'' shows
the MemTotal is 1033800 KB if assigning 1G memory for hvm
guest.
>From my test, cpu2k needs no more than 900M memory to run with 4 users. If I
assign 1.2G memory for hvm guest, the free memory is more than what is freed
when assign 1G memory for native while running cpu2k, but it will still fail.
Best regards,
Fan

-----Original Message-----
From: Keir Fraser [mailto:Keir.Fraser@cl.cam.ac.uk] 
Sent: 2006年10月10日 22:20
To: Zhao, Fan; xen-devel@lists.xensource.com
Subject: Re: [Xen-devel] SPEC cpu2000 will always fail for ia32 smp hvm guest

On 10/10/06 15:02, "Zhao, Fan" <fan.zhao@intel.com> wrote:
> Thanks for your comments. Actually, I do not give a swap partition for both
> native and hvm guest when I do the test.
> One point which deserves to mention is that the tests for hvm guest and
xenu
> guest are based on the same image file, which helps to ensure that hvm and
> xenu guests share the same testing environment. At this point, xenu guest
> doesn''t have a swap partition either, but cpu2k could run
successfully with 1G
> memory in it.
> Another point for reference is that cpu2k with 4 users could pass on ia32e
hvm
> guest with 1G memory, which also does not have a swap partition.
If you run ''cat /proc/meminfo'' on native and on HVM, do the
MemTotal figures
match up? If so, this must be an issue within the guest -- it would indicate
that we are providing the guest with the same amount of memory as was
available to it when running on native.

 -- Keir

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

On 10/10/06 16:43, "Zhao, Fan" <fan.zhao@intel.com> wrote:
> If I assign 1G memory for native, ''cat /proc/meminfo''
shows the MemTotal is
> 1034672 KB, while ''cat /proc/meminfo'' shows the MemTotal
is 1033800 KB if
> assigning 1G memory for hvm guest.
> From my test, cpu2k needs no more than 900M memory to run with 4 users. If
I
> assign 1.2G memory for hvm guest, the free memory is more than what is
freed
> when assign 1G memory for native while running cpu2k, but it will still
fail.
I wonder if SPECint tries to work out how much memory it has available to it
in some way, and decides how big to make test-program heaps and so on based
on this. Clearly something must be different in the environment to explain
this difference, since the kernel and application bits are the same in both
cases and HVM has the same or more memory and yet fails. I suspect this one
will remain a mystery for a while. :-)

 -- Keir



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