Re: [libvirt] [RFC][PATCHv1 2/5] libvirt - new API for getting percpu statistics of VM
On Tue, 19 Apr 2011 12:29:05 +0100
"Daniel P. Berrange" <berrange(a)redhat.com> wrote:
On Fri, Apr 15, 2011 at 04:43:03PM +0900, KAMEZAWA Hiroyuki wrote:
> On Fri, 15 Apr 2011 09:43:15 +0200
> Matthias Bolte <matthias.bolte(a)googlemail.com> wrote:
>
> > 2011/4/15 KAMEZAWA Hiroyuki <kamezawa.hiroyu(a)jp.fujitsu.com>:
> > Also do you really need the absolute CPU time?
> yes, for virt-top -1.
I don't think that is correct. What virt-top ultimately displays
to the end user is a % utilization. So it would be happy to get
data in either % util & display that directly, or get absolute
CPU time and calculate deltas to display % util.
virt-top calculate delta by itself.
I wonder if we could provide an API that could be used to provide
either per-VCPU or per-PCPU statistics in one go, so that we
can finally have an API for cpu time that works well for VMWare.
VMWare ?
eg
typedef struct _virDomainCPUTime virDomainCPUTime;
typedef virDomainCPUTime *virDomainCPUTimePtr;
#define VIR_DOMAIN_CPUTIME_ABS "abs"
#define VIR_DOMAIN_CPUTIME_UTIL "util"
...define more stats later if desired...
To calculate 'util', we need 'interval' for calculating it.
What interval do we use ? The libvirt need to calculate it periodically
by making a thread ?
"util" can be very easily calculated by
util = cpuTime-delta / (interval(sec) * 1000000000) * 100.
So, it depens on interval. For example, when we make fair-share-scheduler's
balancing interval to 1sec, we never want interaval smaller than 1 sec.
IIUC, virt-top users can specify 'interval'. So, virt-top needs to
use absolute cpuTime delta even if libvirt provides 'util'. Because to show
what users want, virt-top needs to use interval users want.
struct _virDomainCPUTime {
char field[VIR_DOMAIN_CPU_TIME_FIELD_LENGTH];
unsigned long long value;
};
typedef enum {
VIR_DOMAIN_CPU_TIME_VIRTUAL = 0,
VIR_DOMAIN_CPU_TIME_PHYSICAL = 1,
} virDomainCPUTimeFlags;
int virDomainGetCPUTime(virDomainPtr dom,
virDomainCPUTimePtr stats,
int *nstats,
unsigned int flags);
When querying per-virtual CPU time, nstats would be the number of
virtual CPUs in the guest. eg
Why we take care of vcpus ?
virDomainInfo info;
virDomainGetInfo(dom, &info);
virDomainCPUTimePtr stats;
stats = malloc(sizeof(*stats)*info.nrVirtCPU);
virDomainGetCPUTime(dom, stats, info.nrVirtCPU,
VIR_DOMAIN_CPU_TIME_VIRTUAL);
Or to get the break down per physical CPU, use the VIR_NODEINFO_MAX_CPUS
macro
per vcpu pcpu time is not what I want. Hmm, others needs such information ?
And virt-top need to re-calculate per-vm pcpu time by making sum of cputime per
vcpu on pcus ? Anyway, cpuacct cgroup/ Linux kernel cannot provide per-vcpu-per-pcpu
time and I cannot implement qemu driver.
Thanks,
-Kame