An Overview of Distributed Real-Time Systems Research

By Brian DemersMarch 24, 2003

CS 535, Spring 2003

Project Overview• Examine efforts to combine real-

time and distributed computing ideas

• Summarize trends, focal points– Tried not to filter ideas

• Examine one or more sample implementations (if possible)

• Work in progress!

Presentation Overview

• Definitions of key terms, issues• Discussion of key research areas,

developments– Networks– Scheduling

• To-do list• Conclusions, questions,

suggestions

Definitions

• Hard vs. soft real-time• Distributed systems

– Did not consider traditional parallel computers

General focus: soft RT systems

Networking

Networking

• Inherently non-deterministic– Research focus is on LAN

configurations

• Main approaches– Improve determinism of existing

protocols– Use probabilistic approach

Ethernet

• Pros– Cheap– Widely used

• Cons– Not designed for real-time use– Packet arrival times can vary widely

Ethernet Improvements

• Halmsted University– Hardware solution:

Propose a network connected exclusively by switches (eliminating collisions)

– Switches perform data buffering, manage real-time streams

Switch

Ethernet Improvements (cont.)• Traffic smoothing

(Kweon, Shin, and Workman)– Probabilistically limit

traffic into network

• Virtual token ring (Chiueh and Venkatramani)– Switch to token-based

scheme when RT traffic detected

Data-link Layer

Regulator

Higher NetworkLayers

Traffic smoothing approach

Token-Based Protocols

• Pros– Fairly deterministic (for a network)

• Cons– Not as widely available– Latencies can be high (according to

some)

Token-Based Protocols (cont.)

• FDDI (Fiber Distributed Data Interface)– Uses token-based, ring topology– Data rates equivalent to Fast Ethernet– Fault tolerant

• RTFC (RT Fiber Communications)

• Not much recent work on IEEE 802.5

Scheduling

Scheduling

• Packet scheduling• Task scheduling• Resource scheduling

Packet Scheduling

• Vital for RT communications• Some probabilistic work • Some work with co-scheduling RT

and non-RT packets

Packet Scheduling (cont.)

• Concatenated Hybrid Automatic Repeat Request (Uhlemann et al.)– Repeated transmissions– Signal averaging

• (m, k) scheduling (Hamdaoui and Ramanathan)– Some work on a distance-based priority

scheme to guarantee this– Extended to multi-hop networks

Packet Scheduling (cont.)

• Combining RT traffic with non-RT traffic (Chakraborty, Gries, and Thiele)– Use Earliest Deadline First (EDF)

scheduling– Assign deadlines so that RT & NRT

traffic can coexist– Claimed 25-45% improvement for

NRT traffic

Task Scheduling

• Landmark paper: Liu & Layland (1973)– Analyzed static and dynamic scheduling

(single processor)– Many efforts to extend this

• Distributed Generalized Multiframe (DGMF) model (Chen, Mok, and Baruah)– Presents new analytical framework

Task Scheduling (cont.)

• Breaking tasks into subtasks (Kao and Garcia-Molina)– Setting subtask

deadlines– Methodologies,

performance issues

Task

Subtask

Subtask

Sub-task

Subtask

Task Scheduling (cont.)

• Resource fragmentation (Bestavros)– Load balancing– Potentially bad for

CPU-intensive jobsCPU 1 CPU 2 CPU 3

60%50%

70%

Example:

Sample Systems

Sample DRT Systems

• Everyone cites the same examples– Air-traffic control, factory automation,

nuclear power plants, etc.– Difficult to find much info about these

topics!

• Some work from avionics• Possibly some air-traffic systems

(FAST?)

Summary

Distributed RT: Research Focus

• Networking– Extending Ethernet to RT apps– Packet scheduling issues

• Working with “soft” RT systems– Leads to probabilistic techniques

To-do: Remaining Efforts

• More detail on resource usage, scheduling

• RT evaluation of token-based networks

• Find and examine sample implementation

Questions/Suggestions?