1
Time In Systems
Ashok K. Agrawala
Department of Computer Science
University of Maryland
College park, MD 20742
agrawala@cs.umd.edu
6/27/2016
University of Maryland, College Park
2
View of a System
• Collection of active entities (agents)
– Active Entity
» State
» Communication
• Exchange information with
– other entities
– environment -- non-specific entities
» Actions
• Actuators
• Means for interaction
• Hierarchical Views
– Aggregation
– Refinement
• Observability
– States/actions/messages visible
• System Dynamics
– Changes over time
6/27/2016
University of Maryland, College Park
3
Real Time and Virtual Time
• Real Time
– Monotonic
– Universal
– Uncontrolled variable
• Virtual Time
– Attached to an activity
– Progresses only when the activity makes a progress
» Gets mapped to real time
• Coordination among activities
– Defined with respect to the virtual time
– Carried out in real time
6/27/2016
University of Maryland, College Park
4
Virtual time for a process
• Virtual Single-Process Machine Abstraction
• The time seen by the process in this
abstraction is the Virtual Time
• Process
– Sequence of steps
» Assign time for each step
• May be assigned as a function - depending on the processor speed
etc.
» Assign Virtual Time to the process
– When can the assignment be done?
» Fixed execution sequence
» Data dependent execution sequence
6/27/2016
University of Maryland, College Park
5
Example - Video Tape
Video tape
– Sequence of frames {n}
– Each frame has to be displayed at a fixed time relative to
the start of the tape
– When we hit start
» assign the real time of start instant to the virtual time
for the first frame of the tape
» Subsequent frames are assigned real time linearly
– Pause
» Real time progresses while virtual time has stopped
– Fast Forward
» Virtual Time runs faster than real time
– Rewind
» Real Time moves forward, virtual time moves
backwards
6/27/2016
University of Maryland, College Park
6
Pause
Play
Pause
Play
Virtual Time of Process i
Real Time
6/27/2016
University of Maryland, College Park
7
Fast Forward of Virtual Time
Play
Fast Forward
Play
Virtual Time of Process i
Real Time
6/27/2016
University of Maryland, College Park
8
Rewind of Virtual Time of Process j
Process j virtual time
Rewind
Undo code operation
Real Time
6/27/2016
University of Maryland, College Park
9
Multi- Process Environment
• Single processor shared among several
processes
• The virtual time of a process only makes
progress when CPU is assigned to it
• The Resource Manager (operating System)
has to be treated as another process which
takes resources/time
6/27/2016
University of Maryland, College Park
10
Mapping of Virtual Time
Process 1 virtual time
Process 2 virtual time
Process 3 virtual time
Real Time
6/27/2016
University of Maryland, College Park
11
Synchronization
• Process synchronization requirements are
usually specified with respect to the virtual
time
• They have to be carried out in real time
• Synchronization mechanisms
– Coordinate the mapping of multiple processes virtual time
to a common real time with the scheduling constraints
6/27/2016
University of Maryland, College Park
12
Synchronization of Multiple
processes
Process 1 virtual time
Process 2 virtual time
Process 3 virtual time
Real Time
6/27/2016
University of Maryland, College Park
13
Virtual Time in Virtual Worlds
• In Virtual Worlds we control all activities
explicitly
• With explicit understanding of virtual time we
can more effectively
manage/control/understand the interactions
6/27/2016
University of Maryland, College Park
14
View of Processing
• Collection of Requests
• Request
– A collection of activities
– Defined constraints
» Static
» Dynamic
• To meet the processing requirements the
requests must be processed meeting all the
constraints
6/27/2016
University of Maryland, College Park
15
Designing a System
• Decomposition
–
–
–
–
Modular Design
Functionality
Resources
Interfaces
• Aggregation
– Coordination
– Composition
» Functionality
» Behavior
• Performance
• Timing
• Interfaces
– Functional -- Temporal
Role of time in system design
University of Maryland, College Park
6/27/2016
16
Absolute Time
•
•
•
•
•
Continuous, real number
Universal
Notion of Present, Past and Future
Spatial locality
Absolute Time at different spatial localities
– Related
» In quantifiable way
– Relativity Theory
6/27/2016
University of Maryland, College Park
17
Physical Time/Local Time
• As used in a physical system
• Related to Absolute Time through CLOCK
– Physical Device
Abs Time
» defined local time
Clock
» observable
Local Time
– Granularity - Tick Time - Finite
– Accuracy is finite
– In a Good clock
» Relationship with Absolute Time is an invariant
6/27/2016
University of Maryland, College Park
18
Knowledge of Time
• Only way of knowing Local Time is by reading
the clock
• Read the clock
– Takes time to read the clock
» If fixed, can factor it
» If variable, limits the accuracy about the knowledge of
time
– Local Time
• Only know that the local time is > time read
from the clock
• If we know the properties of the clock, can
relate the time read to Absolute Time
6/27/2016
University of Maryland, College Park
19
Notion of Time in Systems
• Local Time
– When a system is distributed there may be multiple local
times in a system
• Time Horizon
– Ignore before and after
• Time scale
– Granularity
– Simultaneity
0
Past Horizon
Future Horizon
6/27/2016
University of Maryland, College Park
20
Local Time
•
•
•
•
Specified by a clock
Granularity > granularity of clock
Accuracy of representation
Accuracy of observation
– Depends on the variability of reading latency
6/27/2016
University of Maryland, College Park
21
Time Granularity
Granularity
• Human notions of time
– non-uniformity of time scale
• Machine notions of time
– Uniform time scale
– Models of dynamics
» Differential Equation Models
– Uniform notion of time
– Granularity
» greater than the granularity of the clock used
Time
• Should we use the non-uniform time
granularities in systems?
– Used to some degree today
6/27/2016
University of Maryland, College Park
22
Fusion
• All interactions requires temporal
coordination as well as spatial coordination
– Interactions are physical exchange of information/signals
• The control of interactions is through the use
of local time but takes place in absolute time
• Fusion of local time is required
– Commonality of units, scale, origin, etc.
6/27/2016
University of Maryland, College Park
23
Fusion
• When two active entities have to interact
– Must have common notion of local time
» From the knowledge of TA, A should be able to infer TB
with reasonable accuracy
» Use
• Common clock
• Synchronized clock
• Execute special protocol for coordination
6/27/2016
University of Maryland, College Park
24
Buffers
• Entities which can hold a signal for some time
– Fixed time
– When cleared
• Buffers are the primary means of providing
one directional temporal matching
– A places information in the buffer
» B can pick it up any time after that
• Use of buffers increases latencies and delays
6/27/2016
University of Maryland, College Park
25
Temporal Locality:
Simultaneity
• What A and B consider to be simultaneous
• May be within the granularity of their clocks
• Simultaneity should carry with it the time
scale/granularity
6/27/2016
University of Maryland, College Park
26
Distributed System
• Has multiple localities
• Each locality has a local time
• Distributed system must maintain time fusion
for all localities (temporal consistency)
• Inaccuracies in time fusion
– Longer delays - lower throughputs
– Larger buffers
– etc.
6/27/2016
University of Maryland, College Park
27
Coordination
• Activities of multiple autonomous units have
to be coordinated
– WHY????
• A key to all systems
• Coordination of
– Two
– More than two
6/27/2016
University of Maryland, College Park
28
Modes of Coordination
• Event based
– A causes an event E which is visible to B
• Time based
– Agree to carry out a coordinated action at prearranged
time
6/27/2016
University of Maryland, College Park
29
Event Based Coordination
• Event visible to A and B
– Example - signal sent by A to B
• Signal takes time to propagate from A to B
• B must be monitoring the signal for it to see
– A can not make B to do anything unless B is listening
– If B starts monitoring after A sends the signal, it is lost
– And B must have some temporal coordination before they get in
the event based coordination mode
6/27/2016
University of Maryland, College Park
30
Event Based Coordination
• Most monitoring is done in a loop with finite latency
– There are delays of transfer of signal and the latency of
monitoring cycle
• Handshake protocols may confirm to A that B saw the
signal
– Cause additional latency
• Most software systems used Event based
coordination
• Buffers may be used to account for delayed
detection of signal (temporal Matching)
Temporal variability may continue to increase
when multiple step coordination is required
6/27/2016
University of Maryland, College Park
31
Time Based Coordination
• A and B must have a common notion of local time
– Time Fusion
– Knowledge of drift and drift correction
• Must prearrange the time, TAB, for coordinated
action
– Arrangement must be completed before TAB
» It may be dynamic
• Most hardware systems use time based
coordination
– Clock Based synchronous circuits
• Buffers may be used for temporal matching
6/27/2016
University of Maryland, College Park
32
Coordination
• With the same clock resolution
– Time based coordination can be
» temporally more accurate
» require less overhead of messages and signals
than event based coordination
6/27/2016
University of Maryland, College Park
33
Coordination among Multiple Agents
• The complexity of Event based schemes
increases as the number of agents increases
– Communication medium
» Point to Point
• Many messages- sequential or in parallel
» Broadcast
• Complexity of Time based schemes remains
the same
6/27/2016
University of Maryland, College Park
34
What takes time
• Agent uses ITS resources to carry out an
activity
• Use of resources takes time
6/27/2016
University of Maryland, College Park
35
Coherence of a System
• Coherence
– The way interaction/coordination is managed in time
• Completely Coherent
– All interactions take place at precisely known times
• Incoherent
– All interactions are event driven with no notions of time
mentioned explicitly
• Controlled Coherence
– Time variability of interactions is within defined/know
bounds
6/27/2016
University of Maryland, College Park
36
Coherent Systems
• Time Based Coordination
• Accurate Clock
• Knowledge of global time
– Implicit
– explicit
• Resource Implications
– Need clock interrupt for any impending action
• Brittle
• Rigid/Inflexible
6/27/2016
University of Maryland, College Park
37
Incoherent Systems
• Event Driven Coordination
• No explicit knowledge of time
• Significant variability in temporal
performance
• Overhead
– Buffer space for holding messages
– Detection of message arrival
» interrupt
» Polling
6/27/2016
University of Maryland, College Park
38
Controlled Coherence
The temporal variability for any interaction
is controlled in that the variability does not propagate
• Bounded interval for performing all interactions
• Knowledge of time
– Local with bounded/known variability from global time
• Use knowledge of resource/time requirements and
their bounds to coordinate
• May use time based or event based coordination
mechanisms
6/27/2016
University of Maryland, College Park
39
Resource Requirements
• For every task
» Arrival
» Processing
» Departure
– Resources Required
– Processing Time
» Bounds on variabilities { ta , tb}
– Coordination Requirements
• For every Task
– Define a time before which departure is not permitted
• Consequence is that the variability of the
departure time is bounded
– Trading variability for latency
6/27/2016
University of Maryland, College Park
40
Variability Control
• Departure Time d
• Arrival Time a
• Processing time S
da
a
ds
s
d
6/27/2016
University of Maryland, College Park
41
Example : Cyclone Network
temporal regulator
cyclonode
Non cyclone network
University of Maryland, College Park
Non
cyclone
network
6/27/2016
42
Approach
• Space-time management of all resources
• Periodic operation with fixed period length for
the network
• Fixed chunk size for the network
• Slot
– Time and space description of a chunk at a location at a
time instant
6/27/2016
University of Maryland, College Park
43
Link Model
•
•
•
•
Point to point link
Same slot id for both ends
Timing determined by sending side
Slots managed as circular buffer
6/27/2016
University of Maryland, College Park
44
Cyclonode
controller
• Synchronized local clock
• Calendar based switching operation
• Controller operation
– Clock synchronization
– Control message processing
– Calendar management
6/27/2016
University of Maryland, College Park
45
Cyclonode
incoming
link 1
incoming
link m
next free slot
pointer
switch
controller
outgoing
link 1
outgoing
link n
slot buffer
pointer buffer
free slot list
marker checker
6/27/2016
University of Maryland, College Park
46
Temporal Regulator
• Special cyclonode interfaces hosts
• Buffers absorb host timing variability
• Supports multiple traffic types
6/27/2016
University of Maryland, College Park
47
Temporal Regulator
host
nfsp
incoming
link 1
outgoing
link 1
nfsp
incoming
link m
nfsp
nfsp
controller
outgoing
link n
free slot
slot buffer
pointer buffer
free slot list
marker checker
nfsp next free slot pointer
6/27/2016
University of Maryland, College Park
48
Concluding Remarks
• Controlled Coherence provides an approach
to designing complex systems with
known,bounded temporal variability
• Scalability - support for modular design
• Efficient mechanisms can be designed
• Flexibility increased
• Brittleness reduced
• Local Synchronized Clocks are sufficient
6/27/2016
University of Maryland, College Park