17
A Novel Parallel Traffic Control Mechanism for Cloud Computing Zheng Li, Nenghai Yu, Zhuo Hao MOE-Microsoft Key Laboratory of Multimedia Computing and Communication University of Science and Technology of China

A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

  • Upload
    vudung

  • View
    215

  • Download
    0

Embed Size (px)

Citation preview

Page 1: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

A Novel Parallel Traffic Control Mechanism for Cloud Computing

Zheng Li, Nenghai Yu, Zhuo Hao

MOE-Microsoft Key Laboratory of Multimedia Computing and CommunicationUniversity of Science and Technology of China

Page 2: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Outline

IntroductionWeaknesses of HTB Parallel HTB Experiments

Page 3: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Outline

IntroductionWeaknesses of HTB Parallel HTB Experiments

Page 4: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Traffic Control in Cloud Computing

Control the outbound bandwidthrequire an effective bandwidth managementtraffic scheduler & shaper

Hierarchical Serviceidea of cloud computingdifferent service levelan attempt of customized SLAs on bandwidth

A Contradictiondifferent service levels vs. user experience a possible solution : HTB

Page 5: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Hierarchical Token Bucket

HTBa traffic control algorithmcurrently implemented in Linux kernela module of TC (Traffic Control)

Basic ideabandwidth borrowingmake full use of resourcea solution for the contradictionhierarchical service & better user experience

Page 6: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

CBQ vs. HTBCBQ (AR) HTB (AR/CR)

1Gbps 1Gbps

400Mbps 600Mbps

200Mbps 200Mbps 200Mbps 400Mbps

400/600Mbps 600/800Mbps

200/400Mbps

200/400Mbps

200/600Mbps

400/800Mbps

HTB allows bandwidth borrowing to break AR!

CBQ – Class Based QueueingHTB –Hierarchical Token Bucket

(300 300 200 200)(100 100 400 400)

(200 200 200 200)(100 100 200 400)

[300 300]

[400]

Page 7: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Outline

IntroductionWeaknesses of HTB Parallel HTB Experiments

Page 8: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Weaknesses of HTB

Processing speed500Mbps at mostnot eligible for cloud computing

Reasonsthe inherent limitation of sequential programusage of spin-lock in kernel

Page 9: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Outline

IntroductionWeaknesses of HTB Parallel HTB Experiments

Page 10: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Basic Idea

Lock-free FIFOs based pipeliningport HTB from kernel to user spacebased on multi-core architecturetry to eliminate necessity of using locksreduce concurrencyselectively apply lock-free structuresmake it run in a 1-way 2-stage pipeline fashion

Page 11: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Eliminate Locks Basic 2 operations of HTB: enque & deque Remove htb_activate and htb_deactivate in the 2 operations Critical region is reduced to only the packet queues A tradeoff: using locks but no empty queues

vs. elimate locks to parallelize HTB but might exist empty queues

Page 12: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Lock-free FIFOs Selectively used as the packet queue Eliminate time of lock/unlock operations Make it possible for HTB to run in a pipelined fashion We haven’t adopted the advanced cache-line distance and cache-line

aggregation techniques in [1], because unnecessary

Stage1 Stage2Lock-free FIFO

enque

……

deque

[1] J. Giacomoni, T. Moseley, and M. Vachharajani, “Fastforward for efficient pipeline parallelism: A cache-optimized concurrent lock-free queue”, Proc. of PPoPP’08, New York, NY, USA, February 2008, pp.43-52

Page 13: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Outline

IntroductionWeaknesses of HTB Parallel HTB Experiments

Page 14: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Bandwidth Allocation 2 Scenarios: 1Gbps bandwidth & 2Gbps bandwidth The number of users of Scenario 2 are 2 times of that of Scenario 1 Bandwidth for a user is 0.5Mbps/1Mbps and 2Mbps/12Mbps, for

common service(require low band) and special service(require high band)

Trace files are used in the experiments

… … …

… … … … … … …

1G/1G

125M/650M*8

2.5M/13M*50

0.5M/1M 2M/12M

TOTAL BANDWIDTH

USER GROUP

USER

APPLICATION

Page 15: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Results Exp.1 ~ Exp.4: 1Gbps. Exp.5 ~ Exp.6: 2Gbps Exp.1: all users have traffics. Exp.2: 2/3 of users have traffics Exp.3 ~ Exp.4: 64B pkt len. Exp.3: use parallel HTB, Exp.4: use HTB Exp.5 :all users have traffics. Exp.6: 2/3 of users have traffics

FILE #Packets #Pkt Len.

#Max Len.

#Min Len. #Traffic

File-1 2,397,696 782 1500 64 800File-2 2,397,696 782 1500 64 533File-3 9,765,925 64 64 64 800File-4 4,795,392 782 1500 64 1600File-5 4,795,392 782 1500 64 1067

Exp. #Trace #MPPS #Mbps #Enq. #Deq.1 File-1 1.29 1008 0.39 0.542 File-2 1.29 1006 0.39 0.573 File-3 14.1 941 0.39 0.534 File-3 6.7 427 0.64 1.115 File-4 2.60 2033 0.39 0.546 File-5 2.59 2026 0.39 0.58

Parallel HTB can reach 2Gbps for common packet lengths, 300% improvement of the traditional HTB

Page 16: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments:

Results

Output traffic rate of the total traffic

Output traffic rate of a selected user

Page 17: A Novel Parallel Traffic Control Mechanism for Cloud Computing · (200 200 200 200) (100 100 200 400) [300 300] [400] Outline Introduction Weaknesses of HTB Parallel HTB Experiments: