Nathan/Marko,
On July 24, 2016 at 9:03:16 PM, Nathan Scott (nathans@xxxxxxxxxx) wrote:
> Hi Marko,
> It tends to be block devices and network interfaces that are problematic,
> not CPUs FWIW - these are added and removed more commonly.
We havenât started logging any nodes with this pmda, just starting to figure
out what is useful, but whatever the final choice is, it would be ideal if
pmlogger would automatically pickup any new devices. ÂSo if the pmlogger config
is libvirt.* , and the VMs running at pmlogger start time each have only one
network device and one blockdevice, if a VM is then started with multiple of
each, those start getting logged as well.
>
> > > Also, along the vein of PMIDs, and especially depending on how the PMID
> > > "cluster" identifier is used, the size limits of 10 bits (1024 metrics)
> > > and/or 22 bits (if full cluster space available also) are less appealing
> > > than the convenience of the full 32 bit instance identifier space.
> >
>
> Yeah, I wouldn't just dismiss this aspect either - a max of 70-ish block
> devices will surely be limiting to some people.
Only comment I have here is that we have one use case where we allocate a VM
with all the available resources on the host node. ÂSo in practice currently
that means 128 vCPUs. ÂIn theory, with current hardware, I guess that could be
288 vCPUs without oversubscription.
While we donât oversubscribe CPUs, depending on your workload I suppose you
could multiply the above numbers by 2,3,4... in a (maybe) contrived example.
>
> > Martins, you were already trying out this libvirt PMDA after its
> > initial posting and you have lots of valuable experience and insight
> > with PCP also in general, how do you see these two approaches, which
> > one would be more suitable for you and your use cases?
>
> FWIW, it was primarily from problems that Martins and Joe observed that
> we ended up switching the cgroup metrics over to this compound instance
> name approach.
>
Yeah it was something to do with IDs being reused by different cgroups or
something like that if I recall correclty. We had a lot of cgroups coming and
going.
Martins
|