Answer before revealing the answer: the recall effort is what makes it stick. Then mark “I knew it” or “Review,” and filter on “Review” for next session.
Rappel actif
Restitue de mémoire, sans support.
State the three Popek & Goldberg criteria a VMM must satisfy, in your own words.
- Equivalence / fidelity: a program under the VMM behaves as on bare metal, apart from timing and available resources.
- Resource control / safety: the VMM keeps full control of the hardware; a VM can neither seize resources nor bypass it.
- Efficiency / performance: a statistically dominant fraction of instructions runs directly on the CPU, without VMM intervention.
Emulation meets equivalence but not efficiency, which is why it is not virtualization.
Name the five properties of virtualization from the course and give one consequence of each.
- Binary compatibility: the OS and apps run unchanged (except paravirtualization).
- Interposition: every action goes through the hypervisor, enabling overcommit, encryption, replication.
- Isolation: a fault or load in one VM stays contained, which makes consolidation acceptable.
- Encapsulation: a VM’s whole state is a file, so snapshots and clones are possible.
- Portability: a VM can move between hosts, even live (live migration), the basis of HA and DR.
List the five forms of virtualization, with a one-line goal for each.
- Server: consolidate many under-used physical servers into few busy hosts.
- Desktop (VDI): separate the desktop from the device, run it on servers.
- Application: sandbox an app with its runtime, no local install.
- Network: share one physical fabric among isolated logical networks.
- Storage: pool heterogeneous arrays behind one logical view.
Distingue
Deux notions proches : dis ce qui les sépare vraiment.
Full virtualization, paravirtualization and hardware-assisted: how does the guest reach the hypervisor for a privileged operation in each, and which needs a modified kernel?
- Full: the hypervisor traps and rewrites sensitive instructions on the fly (binary translation). Max compatibility, highest overhead. Guest unchanged.
- Paravirtualization: the guest kernel is modified to issue explicit hypercalls. Very low overhead, but needs a ported kernel (classic Xen).
- Hardware-assisted: the CPU has a guest mode (VT-x, AMD-V); sensitive instructions run natively, the hypervisor steps in only on exits. Today’s default.
Virtualization and emulation both run software for another environment. Which satisfies all three Popek & Goldberg criteria, and why doesn’t the other?
Virtualization targets all three. Emulation (e.g. an ARM CPU simulated on x86) meets equivalence but not efficiency: almost every instruction is interpreted/translated rather than run natively, so the performance criterion fails.
Persistent VDI, non-persistent VDI and session-based (RDSH): compare isolation/personalization and cost/density.
Model Personalization Cost / density Persistent VDI Full, dedicated VM kept per user High storage, 1 VM/user Non-persistent VDI Pooled, reset at logoff, profile tools restore identity Lower, one shared image Session-based (RDSH) Weakest, shared Windows Server Lowest cost, highest density DAS, NAS and SAN: access level (block or file), connectivity and sharing for each.
Access Connectivity Sharing DAS block direct cable (SAS/SATA) single server NAS file LAN, NFS/SMB many heterogeneous clients SAN block dedicated fabric (FC/iSCSI) many hosts, native HA
Explique pourquoi
Le mécanisme derrière la règle.
Why does consolidation raise server utilization from ~10% to 70-80%, and why is isolation what makes it acceptable?
“One server per application” leaves each box mostly idle (~10%). Packing many VMs onto one host fills the spare capacity (70-80%). It is only acceptable because isolation keeps a faulty or overloaded VM from taking down or starving its neighbours: fault isolation, software isolation and performance isolation together.
Why is the hypervisor described as a “new attack surface,” and why is a VM escape so serious?
Every VM runs on top of the hypervisor and trusts it to keep them separated. If an attacker escapes a VM into the hypervisor, or compromises the management plane, they can reach every guest on the host at once. Isolation that protects the VMs also concentrates risk in the layer below them.
Why does VXLAN use a 24-bit VNI instead of relying on VLAN IDs, and what problem does that solve?
A VLAN ID is 12 bits, so only 4094 segments, far too few for a large multi-tenant data center. VXLAN’s 24-bit VNI gives about 16 million segments. It also tunnels L2 over L3 (MAC-in-UDP), so a segment can stretch across a routed fabric instead of being confined to one switched LAN.
Why does encapsulation plus portability enable high availability and disaster recovery?
Encapsulation turns a VM’s entire state into a file; portability lets that file run on another host. Together they mean a VM can be copied, replicated off-site and restarted elsewhere, or live-migrated off a failing host, with no reinstall. That is exactly what HA (restart elsewhere) and DR (recover off-site) need.
Cas pratique
Choisis et justifie avec les critères du cours.
A company runs 40 lightly used physical servers, each one application. Which form of virtualization, and name three advantages and one risk you must plan for.
Server virtualization: consolidate the 40 onto a few hosts with a hypervisor.
Advantages: higher utilization and lower CapEx/OpEx, fast provisioning (template clones), live migration and HA, simpler backup (VMs are files).
Risk to plan for: single point of failure, many workloads now share one host. Mitigate with HA clusters, anti-affinity rules and a tested DR plan.
A hospital wants clinicians to reach the same desktop from any ward terminal, with central patching and no data left on the endpoint. Which desktop model fits, and what two components make the “same desktop anywhere” work?
Non-persistent VDI (fresh clone from a golden image each logon), or persistent if each clinician needs installed apps.
Components: a connection broker authenticates the user and hands back a desktop from the pool; profile management (FSLogix, roaming profiles) re-attaches the user’s identity and data to whatever VM they land on. Data and apps stay in the data center, nothing on the terminal.
Trouve l'erreur
Chaque affirmation contient au moins une faute.
“Paravirtualization runs the guest OS completely unmodified, that’s its advantage over full virtualization.”
It is the opposite: paravirtualization modifies the guest kernel so it issues hypercalls. Full virtualization is the one that runs the guest unmodified (via binary translation). Paravirtualization’s advantage is low overhead, not an unmodified guest.
“A snapshot is a backup of the VM, so I can delete the original datastore safely.”
A snapshot shares blocks with the source: it is a point-in-time reference, not an independent copy. Delete or lose the source and the snapshot is worthless. Snapshots are not backups.
“SDN means the switches decide the routes themselves, in a decentralized way.”
SDN separates the control plane from the data plane and centralizes the decision logic in a controller, which pushes rules down (OpenFlow, P4) to the forwarding hardware. The switches forward; they do not decide.
Classe
Range chaque élément et explique les cas limites.
Classify each by OSI scope/role: VLAN, VPN/SD-WAN, VXLAN overlay, SDN, NFV.
- VLAN: L2 switching, segments one LAN into broadcast domains.
- VPN/SD-WAN: L3/WAN, private encrypted connectivity over shared IP.
- VXLAN: L2 over L3, tunnels virtual segments across a routed fabric.
- SDN: control plane, centralized programmable control split from forwarding.
- NFV: network functions (firewall, LB, router) run as software.
Where does the storage-virtualization layer run in each product: Linux LVM / vSAN; IBM SVC / VPLEX; NetApp FlexArray?
- Host-based (in host OS/hypervisor, that host’s capacity): Linux LVM, VMware vSAN.
- Network-based (in-band appliance in the SAN path): IBM SVC, Dell EMC VPLEX.
- Array-based (one controller virtualizes LUNs from arrays behind it): NetApp FlexArray.
Et si...
Prévois la conséquence.
What if VMs are so easy to create that nobody tracks them (VM sprawl)? What are the consequences?
Forgotten VMs waste capacity and licences, and, worse, they sit unpatched and unmonitored, becoming an attack surface and a liability. Consolidation gains are eaten up, and compliance (per-core/per-VM licensing) drifts out of control. It needs lifecycle governance, not just technology.
What if the single host carrying many consolidated VMs fails, and no HA was planned?
Every workload on it stops at once, the downside of consolidation’s single point of failure. Without an HA cluster to restart the VMs elsewhere, anti-affinity to spread critical pairs, and a tested DR plan, recovery is manual and slow. This is why those three are called mandatory once you consolidate.