RESEARCH
My research interests include Real-Time Systems, Real-Time Scheduling,
Fault Tolerance and Low-Power Computing . The main theme of my dissertation
has been Reward-Based Scheduling , which offers a unified
framework to address these issues simultaneously.
In the following, first the basic concepts of
Reward-Based Real-Time Systems are briefly explained,
along with
Related Models . Then, the main contributions
of my dissertation research are summarized in three sections:
Periodic Reward-Based Scheduling ,
FT-Optimality and
Dynamic Voltage Scaling for Low-Power Real-Time Systems .
In a real-time system, each task (job) should be completed by a deadline. However,
network congestions, dynamic conditions, faults or excessive user demand may
easily lead to an overload state, where it is impossible to finish
all the workload in a timely manner with perfect accuracy. Under
this scenario, resorting to a technique which provides
timely but approximate results is crucial for the robustness
and/or graceful degradation of the system.
In a Reward-Based System, a task is logically decomposed into two parts:
a mandatory part which runs first, producing an approximate
result of acceptable quality; and an optional part which refines
the mandatory part's output. A large number
of today's applications, in fact, can assume this semantic distinction:
An image or movie of imperfect visual quality, a skeleton plan,
an approximate computation of the target's location can first be obtained
by the mandatory part, and later refined by the optional part within the
available computational capacity. While the mandatory part must be
completed by the task's deadline, the optional execution can be aborted
once the deadline arrives.
To quantify the refinement process, a reward function is associated
with the execution of each optional part. The reward function will represent
the characteristics of the application at hand: it is an increasing (or at least
non-decreasing) function of the optional part's computation time.
The behavior of most realistic applications
(e.g. multi-media, image processing, real-time planning)
are best captured by general concave and linear functions.
There are a number of Real-time System Models proposed in the literature
which provision for the timeliness versus precision trade-off.
One of the first such models has been the seminal Imprecise Computation
studies, developed by Jane Liu and her colleagues at UIUC, in the early
and mid 90's.
IRIS (Increased Reward with Increased Service) research at
University of Massachusetts-Amherst
(Dey, Kurose, Towsley) was the first to elaborate on the concave
reward functions, though only for aperiodic tasks.
Finally, Q-RAM (QoS based Resource Allocation Model) was developed at CMU
(Lee, Rajkumar, Siewiorek and Lehoczky) mainly to deal with multi-dimensional
QoS requirements of multi-media services. Their research also provided several
efficient approximation algorithms for the inherently intractable
discrete QoS problem.
In many real-time applications, tasks are invoked periodically, such
as those who receive, process and/or transmit sensor data, audio/video
or control signals. Despite the maturity of periodic scheduling research
for Hard Real-Time Systems, optimal periodic reward-based scheduling
problem was open prior to our work, since the early 90's. The apparent difficulty
of periodic reward-based scheduling stems from the equally important
requirements of meeting the hard deadlines of mandatory parts and scheduling
the optional parts (whose total number of instances during the hyperperiod
can be easily exponential) to maximize the total reward.
In [4] and [10], for general
concave reward functions, we showed:
- The existence of an optimal schedule, where a given task receives
equal amount of optional service time from instance to instance
- How to compute efficiently the optimal service times
- The feasibility and optimality of EDF (Earliest Deadline First)
scheduler when used with these computed service times
We also showed that, the optimality of identical service times per instance
no longer holds for extended models (Fixed Priority Scheduling, deadlines
less than the periods), in addition to the intractability of the problem
once one drops the concavity assumption.
Although the RBS framework allows for timely yet approximate results,
the mandatory parts have still hard deadlines and they should
be completed within their timing constraints even in the presence of
faults. Our work in Fault Tolerant Reward-Based Scheduling
is motivated by the observation that the optional parts in the Reward-Based
models create opportunities for recovery operations in case of the transient
faults which affect the mandatory parts. However, scheduling the optional
parts to merely maximize the total reward may leave no recovery time for
the mandatory part(s) affected by faults. On the other hand, applying a hard
real-time fault tolerance technique may yield unacceptably low reward.
We developed the FT-Optimality (Fault Tolerant Optimality)
framework to provision for faults while accruing as much reward as
possible.
Nevertheless, the various characteristics of the task/fault model
deeply affect the existence of efficient FT-Optimal scheduling
algorithms. The dimensions that we considered in our studies are:
- The existence of precedence relations
- Deadline differences among tasks (Frame-based vs. multiple-deadline systems)
- The type of reward functions
- The recovery technique in use
In [11], we focused
on frame-based systems where tasks share a common, end-to-end deadline and the
reward functions are general concave. We showed the optimality of: a
least commitment strategy for tasks with linear precedence constraints
and, the delaying of all optional parts at the end, for independent tasks. In addition,
we provided an efficient algorithm for the computation of service times which provide
the FT-Optimal schedule.
In [9], we investigated the multiple deadlines case for
independent tasks. For one of the recovery techniques proposed ("Immediate Recovery"),
we developed an optimal algorithm for
linear reward functions; while the other one ("Delayed Recovery")
is shown to be inherently intractable even under modest assumptions.
The contribution of [5] is twofold: First, we
considered and solved the general problem of checking the tolerance to faults
of a given RB schedule, even for multiple faults and for recovery blocks
of (potentially) different lengths per task. Second, we presented an
efficient FT-Optimal solution for the chain of RB tasks with different
deadlines.
With the advent of mobile and wireless computing, power-efficiency has become
one of the paramount issues in Computer Science. Dynamic Voltage Scaling
(DVS) techniques rely on dynamically reducing the supply voltage and the
frequency on-the-fly; resulting in increase in latency and considerable
power savings. Yet, for Real-Time Systems, care must be taken not to violate
any timing constraints.
It is remarkable that the DVS problem, can be proven to be equivalent to the
RBS problem for the static version, where each task is assumed to present
its worst-case workload. We provided the optimal speed assignment
algorithm for tasks with different power consumption characteristics
in [8,3].
Nevertheless, in practice, tasks present frequently a workload which is
far less than the worst-case. Consequently, keeping track and reclaiming
the unused computation time by further slowing down the speed may provide
considerable power savings. To this purpose, we developed the
Dynamic Reclaiming Algorithm . To go one step further, one might adopt
an aggressive approach. In fact, it is possible to have a
spectrum of these techniques depending on the aggressiveness level.
An exposition of all these algorithms, along with their feasibility proofs,
can be found in [7] (for frame-based systems) and in
[1] (for general periodic model).
PUBLICATIONS
-
H. Aydin, R. Melhem, D. Mosse and P.M. Alvarez.
Dynamic and Aggressive Power-Aware Scheduling Techniques for Real-Time
Systems. To appear in the
Proceedings of the 22nd
IEEE Real-time Systems Symposium. London, England, December 2001.
- H. Aydin. Enhancing Performance and
Fault Tolerance in Reward-Based Scheduling. Ph.D.
Dissertation . University of Pittsburgh, August 2001.
-
H. Aydin, R. Melhem, D. Mosse and P.M. Alvarez.
Determining Optimal Processor Speeds for Periodic Real-Time Tasks
with Different Power Characteristics.
Proceedings of the 13th EuroMicro Conference on Real-Time Systems (ECRTS'01) . Delft, Netherlands, June 2001.
-
H. Aydin, R. Melhem, D. Mosse and P. M. Alvarez.
Optimal Reward-Based Scheduling for Periodic Real-Time Tasks.
IEEE Transactions on Computers. Vol. 50, no. 2, pp. 111 - 130,
February 2001.
-
H. Aydin, R. Melhem and D. Mosse.
Optimal Scheduling of Imprecise Computation Tasks in the Presence
of Multiple Faults.
Proceedings of the 8th International Conference on Real-Time
Computing Systems and Applications
(RTCSA'00), S. Korea, December 2000.
-
P. M. Alvarez, H. Aydin, R. Melhem and D. Mosse.
Scheduling Optional Computations in Fault-Tolerant Real-Time
Systems.
Proceedings of the 8th International Conference on Real-Time
Computing Systems and Applications
(RTCSA'00), S. Korea, December 2000.
-
D. Mosse, H. Aydin, B. Childers and R. Melhem.
Compiler-Assisted Dynamic Power-Aware Scheduling for Real-Time Applications.
Workshop on Compilers and Operating Systems for Low-Power (COLP'00),
Philadelphia, PA, October 2000.
-
H. Aydin, R. Melhem and D. Mosse.
Task-level Energy Minimization for Reconfigurable Embedded Systems.
Proceedings of the 4th Annual Workshop on High Performance
Embedded Computing (HPEC'00), Lexington, MA, September 2000.
-
H. Aydin, R. Melhem and D. Mosse.
Tolerating Faults while Maximizing Reward.
Proceedings of the 12th Euromicro Conference on Real-Time Systems
(EuroMicro'00),
Stockholm, June 2000.
-
H. Aydin, R. Melhem, D. Mosse and P.M. Alvarez.
Optimal Reward-Based Scheduling for Periodic Real-Time Tasks.
Proceedings of the 20th IEEE Real-Time Systems Symposium
(RTSS'99), Arizona, December 1999.
-
H. Aydin, R. Melhem and D. Mosse.
Incorporating Error Recovery into the Imprecise Computation Model.
Proceedings of the 7th International Conference on Real-Time
Computing Systems and Applications
(RTCSA'99), Hong Kong, December 1999.
-
H. Aydin, R. Melhem and D. Mosse.
A Polynomial-Time Algorithm to Solve Reward-Based Scheduling Problem,
Technical Report 99-10. Department of Computer Science, University
of Pittsburgh, April 1999.
- H. Aydin, R. Melhem and D. Mosse.
Incorporating Error Recovery into the Imprecise Computation Model.
Technical Report 98-09. Department of Computer Science, University
of Pittsburgh, May 1998.
- H. Aydin.
A hierarchical non-linear planner for simple domains. M.Sc. thesis. Control and
Computer Engineering Department, Istanbul Technical University,
February 1994.