Dynamic Resource Management Architectures and Algorithms for Distributed RealTime Applications

1
Dynamic Resource Management Architectures and
Algorithms for Distributed Real-Time Applications

• Frank Drews
• Center for Intelligent, Distributed, and
  Dependable Systems
• School of Electrical Engineering and Computer
  Science
• Ohio University, Athens, Ohio 45701
• drews_at_ohio.edu

2
Goals

• Enhance real-time and QoS capabilities in grid
  middleware to meet the demands of scientific
  applications

3
Grid Middleware Challenges

• Quality-of-Service (QoS) adaptation under QoS
  constraints
• Need for coordinated end-to-end real-time
  enforcement

4
Outline

• Illustrative example
• Grid middleware requirements
• A generic Grid architecture
• Adaptive resource management models and algorithms

5
Illustrative Applications

• Grid-based real-time medical image retrieval
• NASA on-board satellite systems HART

6
Grid-based Real-Time Medical Image Retrieval

• Distributed medical data sets, including X-ray
  images, CAT scans, etc, distributed across
  different domains (medical research centers,
  hospitals, etc.)
• Medical Researchers submit queries via the
  internet to a Medical Image Retrieval System
  (MIRS) server
• MIRS employs static and dynamic image retrieval
  operations on a subset of database objects that
  allow medical researchers establish structural
  similarities between query objects and database
  objects
• The MIRS sends back the best hits
• A query contains a sample object along with a
  variety of requested QoS parameters and real-time
  timing constraints
• Image quality, image size, various dynamic image
  retrieval parameters, similarity metrics,
  locations of image databases, real-time timing
  deadlines, etc.

7
Grid-based Real-Time Medical Image Retrieval
Problems

• Dynamic (content-based) image retrieval is highly
  complex
• We may need multiple high performance computing
  facilities to distribute the load
• Processing of multiple queries
• User queries may have different priorities
• It would be desirable if users could formulate
  their individual QoS trade-offs
• Timing is difficult to predict
• Retrieval operators can run at various levels of
  QoS, resulting in different time and space
  complexities, and thus different running times on
  the (heterogeneous) computing nodes
• The QoS parameters may need to be changed at
  run-time
• Data is highly distributed data sets vary in
  sizes network transfer times difficult to
  predict

Grid Forum Applications Working Group scenario on
parallel tomography is another example of the
need for real-time capabilities in Grid
middleware.
8
General Utility Model

• Real-time and QoS Requirements for Grid
  Middleware are based on a general utility model
  that involves timing and QoS factors
• For example, a scientist may be willing to
  tolerate a delay in getting results in return for
  increased accuracy of the results
• Or, the scientist may value timeliness above all
  else and be willing to sacrifice the quality of
  the computation to achieve results in a timely
  fashion

9
Real-time and QoS Requirements for Grid Middleware

• Support optimization of utility
• Support end-to-end timing constraints grid
  services must include the capability to reserve
  and deliver server resources and communication
  bandwidth when required
• Support varying levels of QoS Grid services must
  include mechanisms that permit the dynamic
  adjustment of QoS parameters
• Coordinate support in all middleware components

10
Real-time and QoS Requirements for Grid Middleware

• Transparent utility optimization support
• Support utility optimization with basic
  infrastructure provided by the middleware and by
  the end-systems that are involved
• Optimization of utility, including timing
  constraints and QoS adjustments, must fit within
  the architecture and existing interfaces of the
  middleware.

11
Hard vs. Soft Real-Time Constraints

• These criteria assume that the applications have
  timing constraints that are somewhere between the
  classical definitions of hard and soft
• i.e. that there is typically high value to the
  system in meeting timing constraints, but that it
  is not absolutely mandatory.
• This in turn implies that the middleware should
  embody sound real-time resource allocation and
  scheduling techniques.

12
Required Grid Middleware Capabilities

• The notions of real-time and QoS that are
  supported by existing Grid middleware are basic -
  they do not support the optimization of utility
• The existing approaches do not support real-time
  constraint enforcement and QoS adjustment that is
  coordinated both end-to-end, and coordinated
  among necessary middleware components of resource
  allocation, scheduling, and bandwidth management

13
Required Grid Middleware Capabilities

• In particular these important real-time and QoS
  capabilities are required from Grid middleware
• Support for adaptive, distributed resource
  management
• Support for distributed end-to-end scheduling
• Support for network bandwidth management

14
Generic Real-Time Middleware Architecture
15
Example
16
Resource Management Algorithm Development

• Feedback control-based QoS optimization
• Host controller / 1 local resource (DQRAM 1-d)
• Host controller / k shared resources (DQRAM k-d)
• Hierarchical controller architecture
• Robust resource allocation for real-time
  applications which process data at rates that
  vary unpredictably over time

17
Feedback Control based Resource Allocation

• Problem Given an amount of available resources,
  provide on-line control of the QoS settings of
  the tasks so as to optimize the overall system
  utility

18
Related Work

• Burns et al. 2000 The meaning and role of
  value in scheduling flexible real-time systems.
• Humphrey et al. 1997 DQM architecture
• QuO 2001 Quality Objects
• QRAM 1997 Quality-of-Service based Resource
  Allocation Model

19
QRAM

• QRAM - Quality-of-Service based Resource
  Allocation Model
• Uses resource profiles and utility profiles
• Tasks can run at various levels of resource usage
  yielding various levels of quality of service
• Determines an optimal resource allocation that
  maximizes the total system benefit
• Does not consider run-time variations such as
  dynamic task arrivals and completions, changes in
  the resource availability
• Is not tolerant to misspecifications of the
  resource profiles

20
DQRAM 1-d

• Dynamic Quality-of-Service based Resource
  Allocation Model (DQRAM)
• We assume a single resource and a system
  (potentially) consisting of hard, soft, and
  non-real-time applications.
• A controller provides on-line control of the soft
  real-time tasks QoS settings so as to optimize
  the overall system benefit
• Approach is based on discrete control theory we
  close the loop by feeding back the actual
  resource utilization to the controller

21
Goals

• Desired properties of the QoS controller
• Low time complexity
• Analytical performance guarantees
• Stability
• Robust against misspecification of resource
  profiles and utility profiles

22
DQRAM 1-d Feedback Controller
23
DQRAM 1-d Feedback Controller

• The controller aims to always run the tasks in
  states for which the total utility is maximized
• The controller also monitors the actual current
  availability of the resource
• Disturbances in the resource utilization will
  generally lead to an error in the predicted
  resource usage of the tasks
• Controller activations
• - Task arrival
• - Task termination
• - End of each task period

24
Properties of DQRAM 1-d

• State-preserving, incremental, tolerant towards
  misspecifications of resource profiles
• Task arrivals and task terminations can be
  accommodated easily and efficiently
• Misspecification of resource profiles and
  instantaneous peaks in resource usage require
  only incremental changes to the current
  allocation

25
Properties of DQRAM 1-d

• Low-complexity

Version 1
Version 2
26
Properties of DQRAM 1-d

• Stability

optimality points
optimal utility curve
approximation algorithm
utility
amount of resource
27
Properties of DQRAM 1-d

• Analytical performance guarantee dynamic and
  static

optimal utility curve
approximation algorithm
dynamic lower bound
utility
static lower bound
amount of resource
28
DQRAM m-d

• Extension to multiple (k) shared resources

29
DQRAM Implementation

• The DQRAM controller has been integrated into the
  QARMA resource manager

30
DQRAM Implementation

• In addition, we have integrated the DQRAM
  controller into the Linux 2.6 Kernel

31
DQRAM - Hierarchical ControlExample 1

32
DQRAM - Hierarchical ControlExample 2

33
Summary

• Requirements for a real-time grid middleware
• Presentation of a generic real-time grid
  architecture
• Overview of our recent progress in algorithms for
  adaptive resource management

34
Future Work

• Finish up the basic research on algorithms
• Develop real-time grid services