|
On 02/07/14 02:18, Sophie wrote:
On 2014-07-01 14:32, Austin S Hemmelgarn wrote:
On 2014-07-01 08:07, Sophie wrote:
Hi, We have
lots of virtual machines running our Oracle Virtual Manager
setup ( Oracle Linux with Xen) on x86. When I've setup new VMs
I've always assigned CPUs to them instead of sharing them
because all our VMs run RHEL and Oracle 11g. In my opinion
this ensures they have all CPU cycles dedicated without any
chance they'll be starved. Their combined SGA and PGA usually
total 3Gb and I've allocated 4Gb of dedicated RAM. Our DBA
team, who were new to XEN and visualization seem to have a
heightened interest in XEN and have asked me this: ** Why
don't we allocated 32 VCPUS to all virtual machines so that
they can share all resources and when they need CPUs they can
access those that were sitting idle ** Their logic was VCPUs
could be better distributed like this. My question to you is
what do you think? Thanks, Sophie
I think that they have failed to understand that when you don't bind
VCPUs to physical CPUs, xen automatically load-balances the VCPUs
dynamically. Personally, I would keep the same number of VCPUs assigned
to each VM, but not pin them to specific physical CPUs (with the notable
exception of giving Domain-0 pinned VCPUs, and making sure that the VM's
don't run on those physical CPUs). In my experience, this tends to get
better performance with I/O heavy workloads such as most database work.
I would not, however, suggest over-provisioning VCPUs by more than
twice the number of physical CPUs.
From available 64 pCPU assign 4 vCPU to each of 15 VM, (and 4 pCPU
to dom0), so that all pCPU are allocated (an example of what you do
now).
At any one time, 10 pCPU could be idle, this is the same problem
with using 15 physical machines each with a single 4 core CPU.
Maybe better to allocate 6vCPU to each of 15VM (sharing pCPU 4-63),
with 4vCPU to dom0 (dedicated pCPU 0-3). Now instead of 10 idle
pCPU, the really busy VM's can use the spare pCPU's, but no single
VM can steal all available CPU time. Unless of course all of your
VM's are 100% busy at the same time, then you might degrade
performance.
I have seen the argument that it might be better to use the xen cpu
credit scheduler as a more fine grained approach to sharing pCPU's,
but I've never actually used that.
BTW, I collect the value for CPU time from xm list every few
minutes. If CPU time used is 500 seconds in the last 5 minutes, and
I allocated 4 vCPU's, then I see the actual CPU consumption is
500/300 or on average, only 1.7 CPU's was required. Doing this can
help you see which VM's might benefit from more available pCPU's, or
really don't need the number you have already allocated.
If I have 4 VM's that use only 10 seconds CPU time per 5 minutes, I
can easily allocate all 4 to a single pCPU ... etc.
Regards,
Adam
|
