[uug] next cpu for virtualization
dfussell at byu.edu
Wed May 16 13:35:12 MDT 2012
On 05/12/2012 09:59 AM, Thomas S Hatch wrote:
> Something else to remember is that the cpus to use should reflect what the
> virtual machines are doing. I have been in a number of situations where I
> wanted lower clock speed but more cores (read "Buy AMD") for some
> applications and fewer cores that deliver better single core performance
> (read "Buy Intel").
> In the end, I normally want more cores when doing virtualization, otherwise
> I would not generally be virtualizing...
Granted, and general network services don't do a whole lot of floating
point ops. With an AMD machine as large as a 2 socket machine with a
total of 32 integer cores would mean between about 16 and 32 virtual
machines; assuming the memory, disk, and network subsystems can keep up
with that. With an Intel, I'd be limited to half of that.
> Finally, remember to look up the right benchmarks. Aaron cited a great
> article because it tested the actual capabilities of the CPUS in a
> virtualization environment, unlike using a windows desktop benchmark, since
> a windows desktop is not something you generally virtualize...
I did see the Ubuntu virtualization review when I went looking for AMD
on Linux scheduler benchmarks to see if the windows benches were a
result of a poor scheduler; I'll admit I moved on quickly after seeing
the missing results for Xen.
Looking at it again, the normalized data makes it look like the 2
processors are equal, when in fact it's just comparing the performance
of each virtual machine monitor to it's bare-metal platform speed.
Looking at openbenchmarking.org for that test, it looks like the raw
data graphs show AMD (at 8 threads) trailing the intel i7 (at 12
threads) by between 1/4 and 2/3 for everything but compile bench and
Perhaps that's instructive as it has two-thirds the thread count, so
maybe they are equally matched per thread. Except when floats are
involved, then AMD rolls-over and plays dead (as an AMD CPU fan, this
bothers me). However, this is still comparing desktop releases of their
architectures. I'm looking for Xeon vs Opteron comparisons, and
current, real-world experience for common server loads (NFS, Samba,
LDAP, LAMP, DNS, etc)
The only other review I've seen was in November, and expressed doubts as
to if the ESX hypervisor was managing the power levels correctly,
generally found the Intel faster, and a server-based price/performance
ratio of one AMD chip at 5:1 (AMD:Intel)
Does anyone on the list even have 1 of the new Xeon and Opteron chips?
More information about the PLUG