Performance Optimization for Cyber Foraging Network via Dynamic Spectrum Allocation Yang Cao, Shiyong Yang, Tao Jiang, Daiming Qu Email:
[email protected]
Huazhong University of Science & Technology, China 4/15/2011
1 / 23
Contents
1
Introduction Background Motivation Objectives
2
System Architecture Network Elements System Workflow
3
Problem Formulation
4
Feasibility Test
5
Proposed Optimal Solution
6
Simulation Results
7
Conclusions
2 / 23
Background (1/2)
Cyber foraging [Satya01] was proposed to make mobile host (MH), e.g., smart phone, offload resource-demanding tasks to other stronger machines termed as surrogates with the sufficient computing capability through the wireless networks.
3 / 23
Background (2/2) Local Task
Remote Task Mobile Host
Input
S1
S2
S3
S4
Output
Surrogate
Upload Data
Execution S1 Speech Capture
S3 Language Translation
S2 Speech Recognition
S4 Speech Synthesis
Download Data
Task: An operation of a mobile application, e.g., using a language translator to translate speeches newly captured by the phone microphone. A MH task consists of several sub-tasks and is partitioned into one local task executed by the MH and one remote task executed by the surrogate. Remote task: It should be offloaded to a surrogate via wireless link. There are three stages: Uploading Stage (Uplink)/ Executing Stage/ Downloading Stage (Downlink). Task completion time (TCT): The TCT of a remote task consists of uplink, downlink transmission time and the time of task execution by the surrogate, which should not exceed a preset threshold. 4 / 23
Motivation Communication overheads of the cyber foraging network (CFN) are quite severe due to the limited spectrum bandwidth in wireless networks. Some recent schemes [Gu04, Ou08, Satya09] have been proposed to reduce communication overheads for the single MH. However, communication overheads are further aggravated in the CFN, where multi-MHs share the same spectrum to offload tasks.
Deadline
5 / 23
Objectives
To pose the problem of the optimal spectrum resource scheduling for multi-tasks offloaded by MHs, which is very important for the future implement of CFN. However, it has not drawn much attention recently. To propose a system workflow to handle task offloading requests for multi-MHs. To propose an algorithm to test the feasibility of satisfying the task completion time constraint for each remote task simultaneously. To derive an optimal solution of the dynamic spectrum allocation for the CFN.
6 / 23
Objectives
To pose the problem of the optimal spectrum resource scheduling for multi-tasks offloaded by MHs, which is very important for the future implement of CFN. However, it has not drawn much attention recently. To propose a system workflow to handle task offloading requests for multi-MHs. To propose an algorithm to test the feasibility of satisfying the task completion time constraint for each remote task simultaneously. To derive an optimal solution of the dynamic spectrum allocation for the CFN.
6 / 23
Objectives
To pose the problem of the optimal spectrum resource scheduling for multi-tasks offloaded by MHs, which is very important for the future implement of CFN. However, it has not drawn much attention recently. To propose a system workflow to handle task offloading requests for multi-MHs. To propose an algorithm to test the feasibility of satisfying the task completion time constraint for each remote task simultaneously. To derive an optimal solution of the dynamic spectrum allocation for the CFN.
6 / 23
Objectives
To pose the problem of the optimal spectrum resource scheduling for multi-tasks offloaded by MHs, which is very important for the future implement of CFN. However, it has not drawn much attention recently. To propose a system workflow to handle task offloading requests for multi-MHs. To propose an algorithm to test the feasibility of satisfying the task completion time constraint for each remote task simultaneously. To derive an optimal solution of the dynamic spectrum allocation for the CFN.
6 / 23
Contents
1
Introduction Background Motivation Objectives
2
System Architecture Network Elements System Workflow
3
Problem Formulation
4
Feasibility Test
5
Proposed Optimal Solution
6
Simulation Results
7
Conclusions
7 / 23
Network Elements
W
less Wire
ire
d
Cloud
Virtual Surrogates
Access Point
Moblie Hosts
Access point (AP): It operates over a certain wireless channel and serves for MHs. Cloud: It supports scalable computing capability [Armbrust09], which can be allocated and mapped into multiple virtual surrogates with different computing capabilities. Central resource scheduler (CRS): It schedules temporal usages of the wireless channel for multi-MHs and the computing capability allocated to the CFN.
8 / 23
System Workflow tRCD tTOD
Ă
Ă Request Collection
Data Trasmission
Idle
Time
Request collection duration (RCD): During the RCD, the CRS collects the requests of offloading tasks from MHs. Denote tRCD as the length of the RCD which is fixed. When the time for the RCD is expired, The CRS would not accept new requests. Then, the CRS allocates channel usage time to multi-tasks according to their task weight sets and the optimization goal. Task offloading duration (TOD): During the TOD, multi-tasks begin to receive offloading services from the CFN. Denote tTOD as the length of the TOD, which equals to the time for completing all the offloading tasks. The scheduling objective of the CRS is to minimize tTOD with the constraint of the tTCT for each task. 9 / 23
Contents
1
Introduction Background Motivation Objectives
2
System Architecture Network Elements System Workflow
3
Problem Formulation
4
Feasibility Test
5
Proposed Optimal Solution
6
Simulation Results
7
Conclusions
10 / 23
Problem Formulation (1/2)
We define a time slot and a wireless channel combination as a spectrum tile (ST). One ST can only be allocated to one task for its uplink or downlink data transmission. Suppose during the current RCD, M requests of offloading tasks have been collected by the CRS, where M ≥ 1. The task i has a task weight set wi = {ui , qi , di , crented , tdeadline }, i ∈ [1, M ]. i i Suppose that the beginning of the TOD is the slot 1. For the slot k, k ∈ [1, tTOD ], the ST allocation policy is termed as ak . Let if uplink, i, 0, if unused, (3.1) ak = −i, if downlink,
which presents that the CRS allocates the k-th ST to the uplink (positive) or downlink (negative) transmission for the task i. ak = 0 denotes that the k-th ST is not allocated to any task (unused).
11 / 23
Problem Formulation (2/2)
se down The TCT of the task i is expressed as tTCT = tup . Since the tTCT is i i i + ti + ti determined by the ST allocation policy ak , we have
tTCT = max{k : ak = −i}, i i = 1, 2, ..., M.
(3.2)
The time of completing M tasks in the CFN is tTOD = max{tTCT , tTCT , ..., tTCT }. 1 2 M
(3.3)
For each task, tTCT ≤ tdeadline . Therefore, the optimization problem is expressed as i i min tTOD s.t. tTCT ≤ tdeadline , i = 1, 2, ..., M. i i
(3.4)
Given task weight set wi , i = 1, 2, · · · , M , the objective function is to find the optimal ST allocation policy ak , k = 1, 2, ..., tTOD .
12 / 23
Contents
1
Introduction Background Motivation Objectives
2
System Architecture Network Elements System Workflow
3
Problem Formulation
4
Feasibility Test
5
Proposed Optimal Solution
6
Simulation Results
7
Conclusions
13 / 23
Feasibility Test Feasibility test: An algorithm to certify whether at least one feasible ST allocation policy exists to satisfy the M TCT constraints simultaneously, or not. The main idea of the proposed algorithm is to from back to front allocate successive unused STs for the data transmission of each task from the ST whose index is equal to the task’s TCT constraint. The algorithm firstly allocates STs to the downlink data transmission of each task one by one with a certain order, and then allocates STs to the uplink data transmission. Spectrum
a1
a2
tMdealineĂ Ă
t2d
t2deadline
t1d t1deadline
Time
Ă
ak
(a) Test for downlink data transmission
Spectrum
a1
tMc
Ă
t1c
Ă
Time
Ă (b) Test for uplink data transmission
ST allocated to downlink data
ST allocated to uplink data
Unused ST
14 / 23
Contents
1
Introduction Background Motivation Objectives
2
System Architecture Network Elements System Workflow
3
Problem Formulation
4
Feasibility Test
5
Proposed Optimal Solution
6
Simulation Results
7
Conclusions
15 / 23
Proposed Optimal Solution The main idea of the proposed method is to adjust the initial feasible ST allocation states to candidate policies, and then pick out the optimal one from all the candidate policies. Suppose there are totally K feasible ST allocation policies under different orders from the feasibility test. In each feasible ST allocation policy, clear the ST allocation for uplink data transmissions while reserve the ST allocation for downlink data transmissions. This operation forms K initial ST allocation states. For a given initial ST allocation state, we propose an algorithm to adjust the initial ST allocation state to the candidate policy of the optimal policy. Spectrum
tMdealine Ă
t1deadline
Ă
Time
Ă
(a) Initial ST allocation state
Spectrum
tMd tMdeadline Ă
U1last
Ă
t1d
t1deadline Time
Ă (b) ST allocation for the uplink data transmission
Spectrum
Ă
U1last
t1d
tMd tMdeadline Ă Ă
t1deadline Time
(c) Adjusting ST allocation for the downlink data transmission
16 / 23
Contents
1
Introduction Background Motivation Objectives
2
System Architecture Network Elements System Workflow
3
Problem Formulation
4
Feasibility Test
5
Proposed Optimal Solution
6
Simulation Results
7
Conclusions
17 / 23
Simulation Results (1/2) We randomly generate the task weight sets. If the generated task weight sets pass the feasibility test, we find the optimal allocation policy through using the proposed algorithm. Four generated task weight sets that have passed the feasibility test:
Task-1 Task-2 Task-3 Task-4
Uplink (STs) 14 6 7 7
Computation (Slots) 7 21 25 19
Downlink (STs) 3 5 5 6
Deadline (Slots) 98 62 45 69
Corresponding optimal ST allocation policy obtained by the proposed algorithm:
1
5
Task 1 Uplink
10
Task 2 Uplink
15
20
Task 3 Uplink
25
Task 4 Uplink
30
Task 1 Downlink
35
40
Task 2 Downlink
45
Task 3 Downlink
50
Time Task 4 Downlink
18 / 23
Simulation Results (2/2) 500 RR EDF Optimal
80
60
40
20
0 100 - 150
150 - 200 200 - 250 TCT Constraint Range (Slots)
(a)
250 - 300
Total Task Completion Time (Slots)
Task Failure Ratio (%)
100
RR
EDF
Optimal
400
300
200
100
0
3 Tasks
4 Tasks
5 Tasks
(b)
We compare the proposed algorithm with two classic task scheduling algorithm, i.e., the round robin (RR) [Sakata71] and the earliest deadline first (EDF) [Doulamis07] algorithms. Scenario (a): The task deadlines (TCT constraints) are generated uniformly within range [100, 150], [150, 200], [200, 250] and [250, 300], respectively. Set M = 3. Define the task failure ratio as the ratio of the failed tasks number to the number of all handled tasks during a certain time duration. Scenario (b): The task deadlines (TCT constraints) are generated uniformly within range [200, 300]. We vary the number of tasks M from 3 to 5. 19 / 23
Contents
1
Introduction Background Motivation Objectives
2
System Architecture Network Elements System Workflow
3
Problem Formulation
4
Feasibility Test
5
Proposed Optimal Solution
6
Simulation Results
7
Conclusions
20 / 23
Conclusions
We formulated and analyzed the optimization problem of the dynamic spectrum allocation. We proposed an algorithm to test the feasibility of satisfying the task completion time constraint for each remote task simultaneously. We derived an optimal solution to the dynamic spectrum allocation. Conducted simulations showed the validity of the proposed dynamic spectrum allocation algorithm.
21 / 23
Thank you! Email:
[email protected]
22 / 23
References 1
M. Satyanarayanan, “Pervasive computing: vision and challenges,” IEEE Personal Communications, vol. 8, no. 4, pp. 10-17, 2001.
2
X. Gu, K. Nahrstedt, A. Messer, I. Greenberg, and D. Milojicic, “Adaptive offloading for pervasive computing,” IEEE Pervasive Computing , vol. 3, no. 3, pp. 66-73, 2004.
3
S. Ou, Y. Wu, K. yang, and B. Zhou, “Performance analysis of fault-tolerant offloading systems for pervasive services in mobile wireless environments,” IEEE ICC , pp. 1856-1860, 2008.
4
M. Satyanarayanan, P. Bahl, R. Caceres, and N. Davies, “The case for VM-based cloudlets in mobile computing,” IEEE Pervasive Computing , vol. 8, no. 4, pp. 14-23, 2009.
5
M. Armbrust, A. Fox, R. Griffith, A. Joseph, R. Katz, A. Konwinski, G. Lee, D. Patterson, A. Rabkin, and I. Stoica, “Above the clouds: A berkeley view of cloud computing,” EECS Department, University of California, Berkeley, Tech. Rep. UCB/EECS-2009-28, 2009.
6
M. Sakata, S. Noguchi and J. Oizumi, “An analysis of the M/G/1 queue under Round-Robin scheduling,” Operations Research, vol. 19, no. 2, pp. 371-385, 1971.
7
N. Doulamis, A. Doulamis, E. Varvarigos, and T. Varvarigou, “Fair scheduling algorithms in grids,” IEEE Transactions on Parallel and Distributed Systems, vol. 18, no. 11, pp. 1630-1648, 2007.
23 / 23