Niraj Tolia
2009-Feb-11 02:24 UTC
[Xen-devel] Impact of P/C-states on server power with xen-unstable
Hi, So, I was playing with C-states (with and without cpuidle on the xen command line) on Intel Xeons and I discovered that turning C-state support on/off makes no difference in terms of the power used by an idle server. The processor in question is a Xeon X5355 and, when cpuidle is enabled, xenpm reports a C1 state and will show that the server is ~99.96% resident in C1 when idle. However, the power consumed by a idle server in C1 is virtually identical to that consumed by the server when C-state support is disabled and the idle server is in either P1 or even P0 (both set manually using the userspace governor and xenpm set-scaling-speed). Is the processor doing something that is not exposed to Xen/xenpm? For what it''s worth, the power used by the server does show a significant difference between 100% busy CPUs at P0 and P1. The relevant ''xm dmesg'' output for the first core is included below. I am running xen-unstable changeset 19104/31983c30c460. Cheers, Niraj (XEN) cpu0 cx acpi info: (XEN) count = 1 (XEN) flags: bm_cntl[0], bm_chk[0], has_cst[0], (XEN) pwr_setup_done[1], bm_rld_set[0] (XEN) states[0]: (XEN) reg.space_id = 0x0 (XEN) reg.bit_width = 0x0 (XEN) reg.bit_offset = 0x0 (XEN) reg.access_size = 0x0 (XEN) reg.address = 0x0 (XEN) type = 1 (XEN) latency = 0 (XEN) power = 0 (XEN) dp(@0x00000000) (XEN) ==cpu0=(XEN) active state: C-1 (XEN) max_cstate: C7 (XEN) states: (XEN) C1: type[C1] latency[000] usage[00000000] duration[0] (XEN) C0: usage[00000000] duration[5708043748] (XEN) xen_pminfo: @acpi_cpufreq_cpu_init,SYSTEM IO addr space (XEN) CPU 0 initialization completed _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Tian, Kevin
2009-Feb-11 02:40 UTC
RE: [Xen-devel] Impact of P/C-states on server power with xen-unstable
Hi, Niraj
C1, i.e ''hlt'', is always enabled regardless of cpuidle
cmdline option which is about other states deeper than C1. In your box, only C0
and C1 are available as reported and thus it''s reasonable for you to
observe no difference here. Try other box with more C-states available. :-)
Thanks,
Kevin
________________________________
From: xen-devel-bounces@lists.xensource.com
[mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of Niraj Tolia
Sent: Wednesday, February 11, 2009 10:24 AM
To: xen-devel
Subject: [Xen-devel] Impact of P/C-states on server power with xen-unstable
Hi,
So, I was playing with C-states (with and without cpuidle on the xen command
line) on Intel Xeons and I discovered that turning C-state support on/off makes
no difference in terms of the power used by an idle server.
The processor in question is a Xeon X5355 and, when cpuidle is enabled, xenpm
reports a C1 state and will show that the server is ~99.96% resident in C1 when
idle. However, the power consumed by a idle server in C1 is virtually identical
to that consumed by the server when C-state support is disabled and the idle
server is in either P1 or even P0 (both set manually using the userspace
governor and xenpm set-scaling-speed). Is the processor doing something that is
not exposed to Xen/xenpm? For what it''s worth, the power used by the
server does show a significant difference between 100% busy CPUs at P0 and P1.
The relevant ''xm dmesg'' output for the first core is included
below. I am running xen-unstable changeset 19104/31983c30c460.
Cheers,
Niraj
(XEN) cpu0 cx acpi info:
(XEN) count = 1
(XEN) flags: bm_cntl[0], bm_chk[0], has_cst[0],
(XEN) pwr_setup_done[1], bm_rld_set[0]
(XEN) states[0]:
(XEN) reg.space_id = 0x0
(XEN) reg.bit_width = 0x0
(XEN) reg.bit_offset = 0x0
(XEN) reg.access_size = 0x0
(XEN) reg.address = 0x0
(XEN) type = 1
(XEN) latency = 0
(XEN) power = 0
(XEN) dp(@0x00000000)
(XEN) ==cpu0=(XEN) active state: C-1
(XEN) max_cstate: C7
(XEN) states:
(XEN) C1: type[C1] latency[000] usage[00000000] duration[0]
(XEN) C0: usage[00000000] duration[5708043748]
(XEN) xen_pminfo: @acpi_cpufreq_cpu_init,SYSTEM IO addr space
(XEN) CPU 0 initialization completed
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel
Niraj Tolia
2009-Feb-11 05:15 UTC
Re: [Xen-devel] Impact of P/C-states on server power with xen-unstable
On Tue, Feb 10, 2009 at 6:40 PM, Tian, Kevin <kevin.tian@intel.com> wrote:> Hi, Niraj > C1, i.e ''hlt'', is always enabled regardless of cpuidle cmdline option > which is about other states deeper than C1. In your box, only C0 and C1 are > available as reported and thus it''s reasonable for you to observe no > difference here. Try other box with more C-states available. :-) > >Ah. I was working under the assumption that C1 not showing up in xenpm''s output meant that hlt wasn''t being used but your explanation makes sense. Let me see if I can find other machines to experiment with. Cheers, Niraj> Thanks, > Kevin > > ------------------------------ > *From:* xen-devel-bounces@lists.xensource.com [mailto: > xen-devel-bounces@lists.xensource.com] *On Behalf Of *Niraj Tolia > *Sent:* Wednesday, February 11, 2009 10:24 AM > *To:* xen-devel > *Subject:* [Xen-devel] Impact of P/C-states on server power with > xen-unstable > > Hi, > > So, I was playing with C-states (with and without cpuidle on the xen > command line) on Intel Xeons and I discovered that turning C-state support > on/off makes no difference in terms of the power used by an idle server. > > The processor in question is a Xeon X5355 and, when cpuidle is enabled, > xenpm reports a C1 state and will show that the server is ~99.96% resident > in C1 when idle. However, the power consumed by a idle server in C1 is > virtually identical to that consumed by the server when C-state support is > disabled and the idle server is in either P1 or even P0 (both set manually > using the userspace governor and xenpm set-scaling-speed). Is the processor > doing something that is not exposed to Xen/xenpm? For what it''s worth, the > power used by the server does show a significant difference between 100% > busy CPUs at P0 and P1. > > The relevant ''xm dmesg'' output for the first core is included below. I am > running xen-unstable changeset 19104/31983c30c460. > > Cheers, > Niraj > > > (XEN) cpu0 cx acpi info: > (XEN) count = 1 > (XEN) flags: bm_cntl[0], bm_chk[0], has_cst[0], > (XEN) pwr_setup_done[1], bm_rld_set[0] > (XEN) states[0]: > (XEN) reg.space_id = 0x0 > (XEN) reg.bit_width = 0x0 > (XEN) reg.bit_offset = 0x0 > (XEN) reg.access_size = 0x0 > (XEN) reg.address = 0x0 > (XEN) type = 1 > (XEN) latency = 0 > (XEN) power = 0 > (XEN) dp(@0x00000000) > (XEN) ==cpu0=> (XEN) active state: C-1 > (XEN) max_cstate: C7 > (XEN) states: > (XEN) C1: type[C1] latency[000] usage[00000000] duration[0] > (XEN) C0: usage[00000000] duration[5708043748] > (XEN) xen_pminfo: @acpi_cpufreq_cpu_init,SYSTEM IO addr space > (XEN) CPU 0 initialization completed > > > _______________________________________________ > Xen-devel mailing list > Xen-devel@lists.xensource.com > http://lists.xensource.com/xen-devel > >_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel