Architecture for Resource Allocation Services Supporting Interactive Remote Desktop Sessions in Util

1
Architecture for Resource Allocation Services
Supporting Interactive Remote Desktop Sessions in
Utility Grids

• Vanish Talwar, HP Labs
• Bikash Agarwalla, Georgia Tech
• Sujoy Basu, HP Labs
• Raj Kumar, HP Labs
• Klara Nahrstedt, UIUC

2
Utility Grids
Data Center
Batch Jobs
Data Center
UTILITY GRID
Interactive Remote Desktop Sessions

• Consolidated data centers
• Blade Servers host all applications
• Data is accessed and stored at consolidated
• storage systems
• Multiple application domains
• Traditional batch jobs
• Interactive Remote Desktop Sessions
• eg. Microsoft Terminal Servers, Citrix, VNC

Thin Clients
3
Utility Grids (contd.)

• Key Characteristics
• On-demand allocation of blade servers
• Sharing of blade servers among multiple users
• Heterogeneity of resources
• Scale of resources, users, applications
• Advantages
• Ease of manageability
• Reduced costs
• Applications of interest
• Batch Technical applications
• Interactive sessions CAD/CAM, financial, office
  applications

4
The Problem

• Design a Resource Allocation Service for
  Utility Grids so as to
• satisfy QoS performance needs of interactive
  remote desktop sessions
• allow for sharing of resources across both batch
  and interactive sessions
• while
• maintaining a high overall throughput of the
  system
• minimizing the wait time for the user requests

5
Outline

• Introduction
• Proposed Solution
• Key Research Contributions
• Proposed Framework
• Solution to Key Research Questions
• Simulation
• Related Work
• Conclusions and Future Work

6
Key Research Contributions (Summary)

• Architectural Guidelines
• for a resource allocation service supporting
  interactive remote desktop sessions
• Simulation Studies
• for mixed workloads in Utility Grids

7
Users Request and Interaction with the System
2. Allocate a compute node for the requested
remote desktop session
8
Proposed Framework
S T A N D A R D I N T E R F A C E
REPOSITORY
Generate Remote Desktop Session Performance Model

• Resource Model
• Application
• Performance
• Models
• 3. Real Time Utilization
• data from compute
• nodes

Input Queue
Site Admission Control
Pending Queue
Resource Assignment
RESOURCE MANAGEMENT SERVER
9
Performance Models

• Application Performance Model
• captures the resource consumption requirement per
  application
• Profile vector for an application Ai represented
  by
• Ai Ci, Ni, Si, LNi, LSi
• Ci CPU utilization requirement
• Ni Network bandwidth requirement (between
  allocated blade server and thin client)
• Si Storage bandwidth requirement (between
  allocated blade server and remote storage
  servers)
• LNi Network latency requirement (between
  allocated blade server and thin client)
• LSi Storage latency requirement (between
  allocated blade server and remote storage servers)

10
Performance Models (contd.)

• Remote desktop session performance model
• Generated dynamically for the set of applications
  requested within a remote desktop session
• Function of individual resource requirements of
  the requested applications
• Multiple possible execution models

Simultaneous Resource requirement taken as sum
of individual application requirements
Sequential Resource requirement taken as max
of individual application requirements
Mixed Resource requirement taken in
between the 2 extremes of simultaneous and
sequential execution
11
Site Admission Control

• Problem
• Determine the set of blade servers that can admit
  the requested remote desktop session
• Input
• Profile for the requested remote desktop session
• Current utilization of the available blade
  servers
• Policies
• Admission Criterion
• Does the available resource utilization on the
  blade satisfy the requested resource utilization
  values for the remote desktop session without
  violating existing allocations?

12
Resource Assignment System
1. Determine available CPU cycles, N/W BW,
Storage BW 2. Determine delta values between
available resources and the desired resources
for the requested remote desktop session 3.
Determine, Wc f(Cdelta, Compute
Intensiveness) WN f(Ndelta, Average
expected display data size) WS
f(Sdelta, Data intensiveness) WNL
f(NLdelta, Interactiveness) WSL f(NSdelta,
Data intensiveness) 4. Weffective Wc WN
WS WNL WSL

• Multi Variable Best Fit Algorithm
• Tight Bin Packing
• to avoid fragmentation
• so that time to wait in pending queue reduces
• Weighted Best fit algorithm
• along dimensions of CPU, network, storage
  utilization

13
Session Admission Control

• Exists at the compute node for every executing
  remote desktop session
• Problem Determine if the remote desktop session
  can admit the requested application
• Input
• Resources allocated to the remote desktop session
• Profiles for the currently executing applications
• Profile for the requested application
• Admission Criterion
• Does the available resource utilization for the
  remote desktop session satisfy the requested
  resource utilization values for the application
  without violating existing allocations?

14
Outline

• Introduction
• Proposed Solution
• Key Research Contributions
• Proposed Framework
• Solution to Key Research Questions
• Simulation
• Related Work
• Conclusions and Future Work

15
Simulation Setup

• Request classification
• Heavy and Light Interactive Sessions
• Heavy and Light Batch Jobs
• Request description
• Day Time Experiment
• Heavy Interactive Session, Light Batch Job
• Night Time Experiment
• Heavy Batch Job, Light Interactive Session

16
Results

• Day Time experiment
• Heavy Interactive Session, Light Batch job

No Resource Sharing among mixed workloads (100
dedicated nodes)
Complete Resource Sharing among mixed
workloads (100 shared nodes)
17
Results (contd.)

• Night Time experiment
• Light Interactive Session, Heavy Batch job

No Resource Sharing among mixed workloads (100
dedicated nodes)
Complete Resource Sharing among mixed
workloads (100 shared nodes)
18
Outline

• Introduction
• Proposed Solution
• Key Research Contributions
• Proposed Framework
• Solution to Key Research Questions
• Simulation
• Related Work
• Conclusions and Future Work

19
Related Work

• Resource Allocation for different application
  domains
• Batch jobs
• 3-tier applications
• Thin Client Systems
• Citrix
• HP CCI
• Clearcube

20
Summary

• Presented a resource allocation architecture
• supports interactive remote desktop sessions
• considers QoS requirements of applications and
  remote desktop sessions
• Simulation studies
• Mixed batch jobs and interactive sessions
• Studied effect on throughput and wait time

21
Future Work

• More simulation studies
• Different resource sharing strategies
• Evaluating tradeoffs among various weight
  assignments
• Evaluation of overall system utilization under
  various workloads and different resource sharing
  strategies
• Architectural enhancements
• Admission Control strategies
• Resource Assignment strategies

22
Questions?

• Thank You !
• Bikash Agarwalla
• bikash_at_cc.gatech.edu
• http//www.cc.gatech.edu/bikash