A framework for embedding realtime adaptability and QoS in nonrealtime OSes

1
A framework for embedding real-time adaptability
and QoS in non-real-time OSes

• Dimitris Nikolopoulos, CSRD
• 3/23/2001

2
Motivation

• Run efficiently real-time jobs on a desktop PC
  connected to the network
• Multiplex real-time jobs with everyday desktop
  activity jobs
• Conflicting requirements
• turnaround time
• throughput
• real-time (deadlines, guarantees)

3
Real-time job requirements

• Resource availability guarantees
• 1 ms. out of each 10 ms. of CPU time
• continuous or quantized
• resources other than CPU subsumed
• Deadlines
• Coping with unpredictable latencies
• memory hierarchy
• system overhead (context switches)
• network latency

4
Real-time OS features

• CPU reservations
• Time constraints
• Sophisticated scheduling policies
• EDF
• Preemptible kernels
• Dedicated real-time mode

5
Non-real-time OS features

• Non-preemptible kernels
• kernel control paths are not dedicated to service
  one job at a time
• Generic time-sharing
• Job classification orthogonal to real-time
• interactive
• batch
• CPU-bound
• I/O bound
• No resource availability guarantees
• No deadlines, load-dependent scheduling epochs

6
Existing real-time extensions to non-real-time
OSes

• Real-time priority class (POSIX standard)
• all jobs declared as real-time run at higher
  priority compared to non-real-time jobs
• Sufficient for simple real-time tasks (e.g.
  playing an MP3 file)
• No guarantees or support for deadlines
• No efficient multiplexing between real-time and
  non real-time
• likely to underutilize processors
• no prioritization between real-time threads

7
Open issues

• How to integrate true real-time features in a
  non-real-time OS ?
• How to achieve efficient symbiosis of real-time
  and non-real-time jobs ?
• How to cope with unpredictable latencies ?

8
Proposed architecture

• A framework for resource negotiation
• CPU
• memory
• bandwidth
• Mechanisms for application-controlled resource
  management
• application-level runtime libraries
• Global resource management in the OS
• Reduced overhead for critical OS paths
• context switches
• buffering

9
(No Transcript)
10
Architectural properties

• Resource negotiation through loads and stores in
  shared memory
• almost predictable latency
• Application-level resource management
• sophisticated runtime systems
• burden on the library rather than the OS writer
• adaptability
• Fast critical kernel control paths
• resource negotiation without context switches
• resuming preempted OS contexts at user-level
• obtaining incoming network data at user-level

11
Advantages

• Performance
• response time
• guarantees
• throughput
• Extensibility
• better resource management policies at the
  application level
• simpler OS extensions
• Integration, deployment
• extensions to a widely used OS (Linux) rather
  than specialized real-time OSes

12
Current prototypes

• Multithreading runtime systems
• fast, non-blocking application-level thread
  management on top of kernel-level EVs
• CPU-intensive parallel jobs, Java
• NUMA memory management
• dynamic data forwarding on DSM systems
• TCP/IP networking
• isockets
• Platforms
• Linux (ix86-based SMPs), IRIX (ccNUMA)

13
Work plan

• Resource negotiation prototypes in Linux
• CPU, memory and bandwidth request/allocation/revoc
  ation
• Introduction of reservations and deadlines
• kernel extensions for resource guarantees
• scheduling plans
• extending the notion of time-sharing epochs
• Evaluation with simple microbenchmarks
• ping-pong client-server apps.
• synthetic reservation benchmarks

14
Some related work

• Nanothreads and related projects at CSRD
• Rialto operating system at Microsoft
• Nemesis operating system at Cambridge
• Exokernels at MIT
• RTX kernel and related products
• Linux kernel development