Re: [PATCH 4/4] tools: add total/local memory bandwith monitoring

[Chao Peng <chao.p.peng@linux.intel.com>]

On Tue, Jan 06, 2015 at 10:29:18AM +0000, Andrew Cooper wrote:
> On 06/01/15 10:09, Chao Peng wrote:
> > On Mon, Jan 05, 2015 at 12:39:42PM +0000, Wei Liu wrote:
> >> On Tue, Dec 23, 2014 at 04:54:39PM +0800, Chao Peng wrote:
> >> [...]
> >>> +static int libxl__psr_cmt_get_mem_bandwidth(libxl__gc *gc, uint32_t domid,
> >>> +    xc_psr_cmt_type type, uint32_t socketid, uint32_t *bandwidth)
> >>> +{
> >>> +    uint64_t sample1, sample2;
> >>> +    uint32_t upscaling_factor;
> >>> +    int rc;
> >>> +
> >>> +    rc = libxl__psr_cmt_get_l3_monitoring_data(gc, domid,
> >>> +                    type, socketid, &sample1);
> >>> +    if (rc < 0)
> >>> +        return ERROR_FAIL;
> >>> +
> >>> +    usleep(10000);
> >>> +
> >>> +    rc = libxl__psr_cmt_get_l3_monitoring_data(gc, domid,
> >>> +                    type, socketid, &sample2);
> >>> +    if (rc < 0)
> >>> +       return ERROR_FAIL;
> >>> +
> >>> +    if (sample2 < sample1) {
> >>> +         LOGE(ERROR, "event counter overflowed between two samplings");
> >>> +         return ERROR_FAIL;
> >>> +    }
> >>> +
> >> What's the likelihood of counter overflows? Can we handle this more
> >> gracefully? Say, retry (with maximum retry cap) when counter overflows?
> > The likelihood is very small here. Hardware guarantees the counter will
> > not overflow in one second even under maximum platform bandwidth conditions.
> > And we only sleep 0.01 second here. 
> >
> > I'd like to adopt your suggestion to retry another time once that happens.
> > But only one retry and it should correct the overflow.
> >
> > Thanks,
> > Chao
> 
> You have no possible way of guaranteeing that the actual elapsed time
> between the two samples is less than 1 second.  On a very heavily loaded
> system, even regular task scheduling could cause an actual elapsed time
> of more than one second in that snippet of code.
> 
On further thought, this could not be right if implemented this only in
tool stack, due to the fact that the duration between two samples can’t
be guaranteed. Even got sample2 > sample1 here, the data may still wrong
as the hardware counter may overflowed more than one times during this
period.

What the hardware guaranteed here is that at most 1 overflow can happen
(which can be corrected by software) when the duration between two samples
is less than 1 second. So only the data that got from two samples which
duration is actually less than 1 second is valid.

The duration must be checked to use the data, this means something must
be done in hypervisor.

My initial solution is: Add a new hypercall to get both the counter
value and the timestamp at that moment(The two operations should be
atomic).
(Looks like not good to add this to existed resource_op hypercall)


Any suggestions?

Thanks,
Chao
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel

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