Disco: Running Commodity Operating Systems on Scalable Multiprocessors

1
Disco Running Commodity Operating Systems on
Scalable Multiprocessors

• E. Bugnion, S Devine, and M Rosenblum
• Stanford University
• Presented by Aaron J Beach

2
Disco The main idea

• Decouple Operating System from machine hardware.
• OS runs on virtual machine
• Virtual machine is assigned resources by Disco
  which manages a pool of processing
  elements/memory resources

3
Goals and Challenges

• Run operating systems efficiently on scalable
  multi-processor systems
• Reduce overhead associated with Disco layer
• Scale above and below the Disco layer
  (i.e. Many VMs and PEs)
• Do this with a small implementation effort that
  does not require major changes to the OS

4
The Problem

• Innovative Hardware(scalable shared memory
  multiprocessors
• Requires significant changes to system software
  to support hardware advantages(i.e. Partitioning
  OS up into units and others)
• Since modern system software is so large this
  requires large amounts of development time
• Challenges- Overhead (memory,processing,replicati
  on)- Resource Management (idle loop)-
  Sharing/Communication

5
Virtual Machines

• Run operating systems efficiently on scalable
  multi-processor systems
• Reduce overhead associated with Disco layer
• Scale above and below the Disco layer
  (i.e. Many VMs and PEs)
• Do this with a small implementation effort that
  does not require major changes to the OS

6
Implementation

• Managing Resources- CPUs- Memory- Advanced
  Hardware (NUMA)- I/O devices- Network
  interfaces
• Changes to OS

7
CPU Management

• Schedules virtual machine/CPU as task
• Sets registers to virtual machine registersand
  runs the task directly
• Controlled (supervised) access to memory

8
Memory Management

• Another level of address translation
• Gives each virtual machine a memory space
  starting at zero in the proper address size
• Uses TLB for physical-to-machine translation
• Stamps entries with address space identifier to
  avoid flushing TLB when changing address spaces

9
Handling Advanced Hardware

• Must deal with possibly unconstant memory
  accesses
• Maintains locality between cache and memory page

10
Managing I/O device access

• First translates into correct address
• Then gives the device exclusively to the machine
• This makes it so that the device does not have to
  actually be virtualized for each machine, simply
  access must be managed

11
Managing Network Interfaces

• Goals- avoid duplication of data if possible
• How - sharing
• Data loaded into the cache can be accessed by
  other Machines, enabling communication between
  machines.

12
Changes to OS and lower system

• Minimal changes, mostly to hardware abstraction
  level (HAL)... very funny
• Had to map machine address into supervisor
  segment (changing header files)
• OS downcalls resource information to Disco
• Specialized OS SPLASHOSvery basic OS supporting
  SPLASH-2 appsmade with Disco in mind

13
Experiments

• Two main things we are interested in- Overhead-
  Scalability
• Simulated on SimOS because FLASH system is not
  yet ready
• Overhead- Compare with the same OS running
  directly on the hardware (memory CPU slowdown)
• Scalability- Measure running with different of
  processors and virtual machines

14
Experiments Overhead - CPU
15
Experiments Overhead - Memory
16
Experiments - Scalability
17
Dynamic Page Migration/Replication

• To achieve locality, hot pages are replicated and
  migrated to satisfy cache misses locally

18
Overview and Conclusions

• Two approaches to scalable shared memory- Large
  OS development- Statically partition and run
  multiple OS
• Disco is more flexible and takes less
  implementation to support a new OS
• Disco is like unimaginitive microkernel(just
  mirrors interface of hardware)
• Low overhead
• Scalability
• Supports advanced hardware
• Small development effort modifying OS