Virtualization Alessandra Scicchitano, SWITCH EGI TF, Madrid - - PowerPoint PPT Presentation

virtualization
SMART_READER_LITE
LIVE PREVIEW

Virtualization Alessandra Scicchitano, SWITCH EGI TF, Madrid - - PowerPoint PPT Presentation

Virtualization Alessandra Scicchitano, SWITCH EGI TF, Madrid September 2013 alessandra.scicchitano@switch.ch connect communicate collaborate Virtualization is so popular because its a wonderful illusion. connect


slide-1
SLIDE 1

connect • communicate • collaborate

Virtualization

Alessandra Scicchitano, SWITCH EGI TF, Madrid September 2013 alessandra.scicchitano@switch.ch

slide-2
SLIDE 2

connect • communicate • collaborate

2

Virtualization is so popular…

…because it’s a wonderful illusion.

slide-3
SLIDE 3

connect • communicate • collaborate

3

Utilization vs Performance

Nothing is perfect, not even magic. The illusion of having more resources doesn’t mean that we have

  • more. It only means that we use more what we have.

Sharing what we have means that sometimes we have to be patient and altruist.

Great performance != altruism

slide-4
SLIDE 4

connect • communicate • collaborate

4

Sharing

u CPU, memory, disk storage, and network access are shared among VMs on the same host. u From a hardware point of view each VM is perfectly isolated. u From a performance point of view…not so much.

slide-5
SLIDE 5

connect • communicate • collaborate

5

Performance in a virtualized environment

Proactive monitoring:

  • CPU Utilization
  • Memory Utilization
  • I/O latency
  • Network

How much can we trust what we see? And how do we handle what we can’t see?

slide-6
SLIDE 6

connect • communicate • collaborate

6

One step back.

  • Resource sharing is managed by the hypervisor.
  • The main task of the hypervisor is to handle all the incoming

requests from the different VMs without creating interference (Is that even possible??)

  • The hypervisor represents basically an extra layer in the system.

This is something to keep in mind

slide-7
SLIDE 7

connect • communicate • collaborate

7

An example: XEN

slide-8
SLIDE 8

connect • communicate • collaborate

8

Xen CPU Scheduler

Credit Scheduler: Ø Based on weight and cap (optional) parameters Ø VCPUs are labeled under if they have existing credits, over otherwise. Ø Every 30ms the priorities (credits) are recalculated Ø Features load balancing of Virtual CPUs across physical ones.

slide-9
SLIDE 9

connect • communicate • collaborate

9

Everything goes fine until…

...the CPUs of the mother host are fully utilized. When the utilization of the underlying system approaches 100%, VMs will have to wait their turns before getting a slice of CPU

slide-10
SLIDE 10

connect • communicate • collaborate

10

Time keeping

The time keeping problem is very well known in virtualized environment (especially for VMWare) Many steps have been done in the direction of a better synchronization of the machines (also during suspension) However time keeping is still an issue when executing performance measurements.

slide-11
SLIDE 11

connect • communicate • collaborate

11

Sharing memory

When the physical memory is 100% used and more virtual memory is required, we can have either:

  • Swapping: This is done indiscriminately and it leads to serious

performance degradation Or

  • Ballooning: It improves resource utilization. It is important not to
  • vercommit too much memory which leads back to swapping.

(This might not be true in all virtual solution. Hyper-V for example doesn’t support overcommitting)

slide-12
SLIDE 12

connect • communicate • collaborate

12

Memory ballooning?

slide-13
SLIDE 13

connect • communicate • collaborate

13

And?

Ok…everything very nice, but can we just go back to the point? Yes. How the hell do we measure performance in a VM? Fair enough!!

slide-14
SLIDE 14

connect • communicate • collaborate

14

Most of the numbers that we see on a monitor running within the VM can be trust, but that it will give us no idea of what happens on the mother host

slide-15
SLIDE 15

connect • communicate • collaborate

15

Something tricky to consider

  • Because of an application a VM needs more

resources

  • It is either stalled (has to wait), or it begins to “steal”

resources from other VMs running on the same host.

  • The other VMs will behave poorly, but they are not

the problem.

slide-16
SLIDE 16

connect • communicate • collaborate

16

The network The network

Virtualization techniques can most easily be broken down into three groups:

  • Emulation
  • Paravirtualization
  • Hardware pass-through
slide-17
SLIDE 17

connect • communicate • collaborate

17

Emulation

slide-18
SLIDE 18

connect • communicate • collaborate

18

Paravirtualization

slide-19
SLIDE 19

connect • communicate • collaborate

19

Hardware pass-through

slide-20
SLIDE 20

connect • communicate • collaborate

20

Paravirtualization is often the solution to boost performance