1. The Whats and Whys of Whole System Virtualization Peter A. Dinda Prescience Lab Department of Computer Science Northwestern University http://plab.cs.northwestern.edu Virtuoso Project: http://virtuoso.cs.northwestern.edu Peter A. Dinda Prescience Lab Department of Computer Science Northwestern University http://plab.cs.northwestern.edu Virtuoso Project: http://virtuoso.cs.northwestern.edu

2. Whole System Virtualization  “Many problems in computer science can be solved by adding a layer of indirection” (bad paraphrase)  Virtualized X  X is a semantically invisible layer of the software stack  Exports exactly the interface it builds on  Adds functionality and/or solves problems  Whole system: span the stack horizontally  “Many problems in computer science can be solved by adding a layer of indirection” (bad paraphrase)  Virtualized X  X is a semantically invisible layer of the software stack  Exports exactly the interface it builds on  Adds functionality and/or solves problems  Whole system: span the stack horizontally

3. OS Virtual Machines  Traditional (Goldberg types I and II)  Run off-the-shelf operating systems  Very low computational overhead but some I/O overheads (arguable how far it can be reduced)  VMware, Microsoft (and VM from the early ‘70s)  Paravirtualized  OS kernels must be ported to them  Very low computational and I/O overhead  Xen, User Mode Linux  Virtual servers  OS kernel extensions (one OS, many instances)  Negligible computational and I/O overhead  Vserver, BSD Jails  Traditional (Goldberg types I and II)  Run off-the-shelf operating systems  Very low computational overhead but some I/O overheads (arguable how far it can be reduced)  VMware, Microsoft (and VM from the early ‘70s)  Paravirtualized  OS kernels must be ported to them  Very low computational and I/O overhead  Xen, User Mode Linux  Virtual servers  OS kernel extensions (one OS, many instances)  Negligible computational and I/O overhead  Vserver, BSD Jails

4. Language Virtual Machines  Compiler targets abstract machine  Usually stack machine  Run-time interprets and dynamically translates to base ISA  Large standard library for I/O  JVM, CLR, (and p-System from late ‘70s)  Arguably also Lisp, Scheme, Perl, Python, …  Compiler targets abstract machine  Usually stack machine  Run-time interprets and dynamically translates to base ISA  Large standard library for I/O  JVM, CLR, (and p-System from late ‘70s)  Arguably also Lisp, Scheme, Perl, Python, …

5. Overlay Networks and P2P  VPNs and VLANs  Multisource multicast (ESM, etc)  Distributed hash tables (Chord, etc)  Resilient routing (RON, etc)  Anonymous routing (Tor, etc)  VM-specialized (VNET, VIOLIN)  VPNs and VLANs  Multisource multicast (ESM, etc)  Distributed hash tables (Chord, etc)  Resilient routing (RON, etc)  Anonymous routing (Tor, etc)  VM-specialized (VNET, VIOLIN)

6. Virtual Storage And Devices  Storage Area Networks  iSCSI  Remote device support  Network block device  Storage Area Networks  iSCSI  Remote device support  Network block device

7. Virtualized Services  Tunneling  ssh  Virtual file systems  System-call interposition  Tunneling  ssh  Virtual file systems  System-call interposition

8. Reducing Complexity  Ownership  Give the user the parallel/distributed systems analogue of a PC  Deployment and distribution  Whole system image  See Potter’s snapshots for a very nice example  Automatic policy avoidance^Wnavigation  Route through the diverse security policies in a multi-site computing environment  Ownership  Give the user the parallel/distributed systems analogue of a PC  Deployment and distribution  Whole system image  See Potter’s snapshots for a very nice example  Automatic policy avoidance^Wnavigation  Route through the diverse security policies in a multi-site computing environment

9. Adaptive Systems  Bring automatic adaptation and resource reservations to existing, unmodified applications  Virtualization as a layer for observation, a provider of adaptation mechanisms, and an impedance matcher to reservations  Bring automatic adaptation and resource reservations to existing, unmodified applications  Virtualization as a layer for observation, a provider of adaptation mechanisms, and an impedance matcher to reservations VM Layer Virtualization Layer Physical Layer

10. Making High-end Computing A Commodity  Virtualization for fungibility  Provider’s perspective  Simple, straightforward abstraction to sell  User’s perspective  Maximum flexibility  “Giant PC”  Virtualization for fungibility  Provider’s perspective  Simple, straightforward abstraction to sell  User’s perspective  Maximum flexibility  “Giant PC”

11. Open-source Virtual Machine Monitor  Type-I OS VMM for modern architectures  Intel’s VT extension to IA32 and IA32e, and AMD’s Pacifica extension to AMD64  Make these commodity architectures virtualizable in the Goldberg sense  VT/Pacifica VMM can be MUCH simpler than existing VMMs for these architectures  Think 50K lines of code (VAX Secure VMM example)  Potentially a very high impact project from this community  Type-I OS VMM for modern architectures  Intel’s VT extension to IA32 and IA32e, and AMD’s Pacifica extension to AMD64  Make these commodity architectures virtualizable in the Goldberg sense  VT/Pacifica VMM can be MUCH simpler than existing VMMs for these architectures  Think 50K lines of code (VAX Secure VMM example)  Potentially a very high impact project from this community

12. Trustless Computing and Language VMs  Trust asymmetry problem in grid and utility computing  Encrypted computation to the rescue  Language VMs are perfect place to implement  Translate binary to binary  Portable  Trust asymmetry problem in grid and utility computing  Encrypted computation to the rescue  Language VMs are perfect place to implement  Translate binary to binary  Portable