1 / 76

Virtualization

Virtualization. Enforced Modularity. Principles of Computer System (2012 Fall). Virtualization. Widely used term involving a layer of indirection top of some abstraction . Virtualization: Multiplexing. Examples: Thread: 1 CPU looks like N CPUs

kirkan
Download Presentation

Virtualization

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Virtualization Enforced Modularity Principles of Computer System (2012 Fall)

  2. Virtualization • Widely used term involving a layer of indirection top of some abstraction

  3. Virtualization: Multiplexing • Examples: • Thread: 1 CPU looks like N CPUs • Virtual Memory: 1 memory looks like N memories • Virtual Circuit: 1 channel looks like N channels • Virtual Machine: 1 machine looks line N machines • Virtual Links – Bounded buffers

  4. Virtualization: Aggregation • Examples: • RAID disks: N disks look like one disk • Channel Bounding: N channels look like one • Compute side: GPUs? ModuleA VirtualizationLayer ModuleA ModuleB ModB ModB ModB

  5. Virtualization: Emulation • Examples: • RAM Disk: Make a RAM memory act like a very fast disk • Apple’s Rosetta Technology: Make a x86 CPU look like a Power PC CPU • Virtual Machines: Virtual card punches, virtual card readers, virtual networks

  6. Virtualization for enforced modularity • Multiple isolated modules on the same computer • Use Operating System abstractions to enforce modularity • Major abstractions: • Memory • Virtual memory address space – OS gives modules their own memory name space • Communication • Bounded buffer virtual link – OS gives modules an inter-module communication channel for passing msgs • Interpreter • Thread – OS gives modules a slice of a CPU

  7. Virtual memory

  8. Operating System • Goals • Multiplexing • Protection • Cooperation • Portable • Performance • Technologies • Virtualization, e.g., memory and CPU • Abstraction, e.g., file, bounded buffer, window

  9. Memory modularity enforcement • No enforcement in physical memory • Any module can overwrite any other • All modules share the same address space (naming context) • Bad things happen Physicalmemorynamespace 0x9000 0x7000 0x4000 0x3000 0x0000 OS gcc disk image access emacs

  10. Virtual memory address spaces • Each module gets its own names • Addresses running from 0 to maximum address • Can’t even name memory of other modules or OS • Error handling • Handles all types of Operating System memory smash errors 2n 0 2n 0 ModuleB ModuleA OperatingSystem Module

  11. Virtualizing Memory • Virtual addresses • Virtual address spaces • Virtual memory manager

  12. Virtual Address • decouple modules with indirection • From a naming point of view • the virtual memory manager creates a name space of virtual addresses on top of a name space of physical addresses • The virtual memory manager’s naming scheme • translates virtual addresses into physical addresses • Managing physical memory transparently • Virtual memory transparently simulate a larger memory than the computer actually possesses

  13. Virtual Address

  14. Virtual Address • The cost of copying the content of a domain, • Complexity and performance • must manage both virtual addresses and a physical address space • must allocate and deallocate them (if the virtual address space isn’t large), • must set up translations between virtual and physical addresses • The translation happens on-the-fly • may slow down memory references.

  15. Translating address using page map

  16. Virtual Address Spaces • provides each application with the illusion • it has a complete address space to itself • The virtual memory manager • gives each virtual address space its own page map • id ←CREATE_ADDRESS_SPACE () • block ← ALLOCATE_BLOCK () • MAP (id, block, page_number, permission) • UNMAP (id, page_number) • FREE_BLOCK (block) • DELETE_ADDRESS_SPACE (id)

  17. Kernel mode • How to protect page tables? • Add one bit to the processor to indicate the mode • change the value of domain registers only in kernel mode • generate an illegal instruction exception otherwise • change the mode bit only in kernel mode • Mode switch • Start in K • K->U: using iret • U->K: interrupt and system call

  18. UNIX Abstractions • Why not using libraries?

  19. Kernel and address spaces • Each address space include a mapping of the kernel into its address space • Only the user-mode bit must be changed when mode switches • The kernel sets the permissions for kernel pages to KERNEL-ONLY • Kernel has its own separate address space • inaccessible to user-level threads • extend the SVC instruction to switch the page map address register

  20. Monolithic kernel • The kernel must be a trusted intermediary for the memory manager hardware, • Many designers also make the kernel the trusted intermediary for all other shared devices • the clock, display, disk • Modules that manage these devices must be part of the kernel program • the window manager, network manager , file manager • Monolithic kernel • most of the operating system runs in kernel mode

  21. Microkernel • We would like to keep the kernel small • Reduce the # of bugs • Restrict errors in the module which generates the error • the file manager module error may overwrite kernel data structures unrelated to the file system • causing unrelated parts of the kernel to fail

  22. Microkernel • System modules run in user mode in their own domain • Microkernel itself implements a minimal set of abstractions • Domains to contain modules • threads to run programs • virtual communication links

  23. Microkernel VS. monolithic kernel • Few microkernel OS • Most widely-used operating systems have a mostly monolithic kernel. • the GNU/Linux operating system • the file and the network service run in kernel mode • the X Window System runs in user mode.

  24. Microkernel VS. monolithic kernel • A critical service doesn’t matter much if it fails • Either in user mode or in kernel mode • The system is unusable. • Some services are shared among many modules • easier to implement these services as part of the kernel program, which is already shared among all modules • The performance of some services is critical • the overhead of SEND and RECEIVE supervisor calls may be too large

  25. Microkernel VS. monolithic kernel • Monolithic systems can enjoy the ease of debugging of microkernel systems • good kernel debugging tools • It may be difficult to reorganize existing kernel programs • There is little incentive to change a kernel program that already works • If the system works and most of the errors have been eradicated • the debugging advantage of microkernel begins to evaporate • the cost of SEND and RECEIVE supervisor calls begins to dominate.

  26. Microkernel VS. monolithic kernel • How to choose • a working system and a better designed, but new system • don’t switch over to the new system unless it is much better • The overhead of switching • learning the new design • re-engineering the old system to use the new design • rediscovering undocumented assumptions • discovering unrealized assumptions

  27. Microkernel VS. monolithic kernel • the uncertainty of the gain of switching • The claims about the better design are speculative • There is little experimental evidence that • microkernel-based systems are more robust than existing monolithic kernels

  28. Exo-kernel

  29. virtual link: bounded buffer

  30. Virtual communication links 2n 0 2n 0 ModuleA ModuleB OS

  31. Enforce modularity for bounded buffer • Put the buffer in a shared kernel domain, the threads • cannot directly write the shared buffer in user mode • must transition to kernel mode to copy messages into the shared buffer • SEND and RECEIVE are supervisor calls • change to kernel mode • copy the message from/to the thread’s domain to/from the shared buffer • Programs running in kernel mode is written carefully • Transitions between user mode and kernel mode can be expensive

  32. Virtual communication links • Bounded buffers interface: buffer ← ALLOCATE_BOUNDED_BUFFER(n) DEALLOCATED_BOUNDED_BUFFER(buffer) SEND(buffer, message) Message ← RECEIVE(buffer) • Producer and consumer problem • Sender must wait when buffer is full • Receiver must wait when buffer is empty • Sequence coordination

  33. Bounded Buffer Send

  34. Bounded Buffer Send/Receive

  35. Bounded Buffer Send/Receive

  36. Send/Receive ImplementationAssumptions • Single producer & Single consumer • Each on own CPU • in and out don’t overflow • read/write coherence • in and out before-or-after atomicity • The result of executing a statement becomes visible to other threads in program order • Compilers cannot optimize it

  37. Multiple Senders • Multiple senders may send at the same time

  38. Race Condition • Race condition • Timing dependent error involving shared state • Whether it happens depends on how threads scheduled

  39. Race conditions ─ Hard to Control • Must make sure all possible schedules are safe • Number of possible schedules permutations is huge • Bad schedules that will and will not work sometimes • They are intermittent • Small timing changes between invocations might result in different behavior which can hide bug • Heisenbugs (Heisenberg)

  40. RaceCondition if two writes p.in=p.in+1 in = 0 time Thread2 Thread1 mov in, %eax(0) add $0x1,%eax mov %eax, in (1) movin,%eax(0) add$0x1,%eax mov%eax,in(1) in = 1 • Anotherpossibleinterleaving: • Everythingworksfine • Debuggingcanbeextremelydifficult

  41. Before-or-After Atomicity • Code assumes that when the producer does: p.in = p.in+ 1 • and the consumer reads in: while (p.in == p.out) do nothing // Spin • will receive the value of in before or after the assignment • Most machines have atomicity when the variable size is less than or equal to the memory width

  42. Before-or-After Atomicity • volatile int in; • Works on 32 and 64 bit machines mov in, %eax add $0x1, %eax mov %eax, in

  43. Before-or-After Atomicity • volatile long long int in; • Works only on 64bit machines, not 32bit mov in, %eax mov in+4, %edx add $0x1, %eax adc $0x0, %edx mov %eax, in mov %edx, in+4

  44. Locks: Before-or-after atomicity • Widely used concept in systems • With a bunch of different names • Database community • Isolation and Isolated actions • Operating systems community • Mutual exclusion (mutex) and Critical sections • Computer architecture community • Atomicity and Atomic actions

  45. Using Locks • Developer must figure out the possible race conditions and insert locks to prevent them • Puts a heavy load on the developer • Developer must place acquire/release of locks in the code • Nothing is automatic • Forget a place and you have race conditions • Forget to release a lock or try to acquire it twice, and you have likely have a deadlock

  46. Send with Locking

  47. Does this work?

  48. Lock Granularity • Systems can be distinguished by number of locks and the amount of share data they protect • Developer must choose a locking scheme to provides the need amount of concurrency • Frequently concurrency leads to complexity

  49. Lock Granularity • Course-grain locking – Few locks protecting more data + Simpler, easier to reason about - Less concurrency • Fine-grain locking – Many locks protecting smaller pieces of data + High amount of concurrency - More complex - Overhead for locks

  50. Simple example of granularity • Allocate a single lock and acquire the lock every time you enter the subsystem and release it when you leave the subsystem • Advantage: Simple. As long as subsystem doesn’t call into itself recursively everything is fine • Disadvantage: No concurrency • Frequently done if a subsystem doesn’t need concurrency while other parts of the system does • Example: Many large software projects • Many modern operating systems (Example: Unix/Linux, NT)

More Related