Download presentation
Presentation is loading. Please wait.
Published byIan Stuart Modified over 11 years ago
1
Bulk-Data Metanet: Virtualization by Example Sergey Gorinsky Applied Research Laboratory Applied Research Laboratory Department of Computer Science and Engineering Department of Computer Science and Engineering Washington University in St. Louis Washington University in St. Louis St. Louis, MO 63130-4899, USA St. Louis, MO 63130-4899, USA 2006/11/8, NSF FIND kickoff meeting
2
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 2 Overview Bulk-Data Transfers Bulk-Data Transfers Bulk-Data Metanet as Part of the Overall Architecture Bulk-Data Metanet as Part of the Overall Architecture Inefficacies of Internet Services for Bulk Data Inefficacies of Internet Services for Bulk Data Potential of Transfer Scheduling Potential of Transfer Scheduling Fruits of Virtualization Fruits of Virtualization Research Agenda Research Agenda
3
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 3 Bulk-Data Transfers Definition Definition transfer time is much longer than round-trip time (RTT) transfer time is much longer than round-trip time (RTT) -deliberately an imprecise definition Performance metric Performance metric transfer time transfer time -and not, e.g., throughput during any smaller time interval Sample applications Sample applications large-scale science large-scale science -e.g., astronomical data sets software download software download -e.g., operating systems
4
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 4 Bulk-Data Metanet as Part of the Overall Architecture Bulk-Data Metanet User Substrate Optical flow switching networkGMPLS networkAnother physical network User Another end-to-end metanet
5
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 5 Inefficacies of Internet Services for Bulk Data In realizing the intended ideal of fair efficient allocation In realizing the intended ideal of fair efficient allocation TCP discrimination against flows with long RTTs TCP discrimination against flows with long RTTs slow convergence of TCP throughput to fair efficient rates slow convergence of TCP throughput to fair efficient rates solutions proposed for the above solutions proposed for the above -numerous congestion control proposals of various efficiency and degree of network support In pursuing a wrong ideal In pursuing a wrong ideal instantaneously fair allocations do not minimize transfer times instantaneously fair allocations do not minimize transfer times prioritized service can improve the average transfer time prioritized service can improve the average transfer time -e.g., Shortest Job First scheduling of CPU -danger: starvation of some (long) transfers
6
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 6 Search for an Ideal Allocation for Bulk-Data Transfers Farsighted Internet congestion control Farsighted Internet congestion control (Infocom 2005 paper by Key, Massoulie, and Vojnovic) (Infocom 2005 paper by Key, Massoulie, and Vojnovic) transmits nothing when congestion is heavy transmits nothing when congestion is heavy transmits more than TCP when congestion is light transmits more than TCP when congestion is light can starve large transfers under persistent congestion can starve large transfers under persistent congestion provides no benefits within the class of bulk-data transfers provides no benefits within the class of bulk-data transfers Isolated specialized networks Isolated specialized networks (UltraScience Net, CHEETAH, OSCARS, DRAGON) (UltraScience Net, CHEETAH, OSCARS, DRAGON) schedule transfers based on message sizes and topology schedule transfers based on message sizes and topology rely on GMPLS to establish dedicated end-to-end channels rely on GMPLS to establish dedicated end-to-end channels are technologically limited to transfers times of minutes or more are technologically limited to transfers times of minutes or more have limited reach and high cost have limited reach and high cost
7
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 7 Potential of Transfer Scheduling: Simulation View Constraints Constraints no transfer finishes later than with maxmin-fair rates no transfer finishes later than with maxmin-fair rates network has a single bottleneck network has a single bottleneck Virtual Finish Time First (ViFi) algorithm Virtual Finish Time First (ViFi) algorithm transfer messages one at a time with preemption in the order of their maxmin-fair finish times transfer messages one at a time with preemption in the order of their maxmin-fair finish times Simulation settings Simulation settings 3000 messages on 10 Tbps path 3000 messages on 10 Tbps path Poisson arrivals with the average rate of 1 message per second Poisson arrivals with the average rate of 1 message per second uniformly distributed message weights from set {1, 2, 3, 4} uniformly distributed message weights from set {1, 2, 3, 4} Pareto-distributed message sizes with Pareto index 1.5 and minimum size 500 GB Pareto-distributed message sizes with Pareto index 1.5 and minimum size 500 GB
8
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 8 Number of Pending Messages (one experiment)
9
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 9 Distributions of Transfer Times (10000 experiments)
10
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 10 Fruits of Virtualization Better services and lower costs for users Better services and lower costs for users specialized service improving transfer times specialized service improving transfer times lower costs than in isolated specialized networks lower costs than in isolated specialized networks New sources of revenue for substrate providers New sources of revenue for substrate providers ability to sell more advanced technology (e.g., optical flow switching) to metanet providers at a higher price despite limited deployment ability to sell more advanced technology (e.g., optical flow switching) to metanet providers at a higher price despite limited deployment Lower costs and new revenues for metanet providers Lower costs and new revenues for metanet providers dynamic lease and release of physical infrastructures dynamic lease and release of physical infrastructures access to various types and sets of end-to-end resources (e.g., both GMPLS and optical flow switching) access to various types and sets of end-to-end resources (e.g., both GMPLS and optical flow switching) ability to attract new types of users (e.g., software downloaders) ability to attract new types of users (e.g., software downloaders)
11
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 11 Research Agenda Algorithms for optimal transfer scheduling Algorithms for optimal transfer scheduling hard problem in general topologies hard problem in general topologies efficient implementations for online operation efficient implementations for online operation Security and robustness Security and robustness human factor human factor enforcing the resource allocation schedule enforcing the resource allocation schedule Efficiency coordination and timing Efficiency coordination and timing distributed activation and renegotiation of the allocation schedule distributed activation and renegotiation of the allocation schedule new algorithms for transmission control at hosts and routers new algorithms for transmission control at hosts and routers Interface between the metanet and physical substrate Interface between the metanet and physical substrate link capacities, processing power, reconfiguration time, buffer space, geographical location? link capacities, processing power, reconfiguration time, buffer space, geographical location?
12
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 12 Additional Slides
13
Sergey Gorinsky (Washington University in St. Louis), Bulk-Data Metanet: Virtualization by Example, 2006/11/8 13 Scheduling versus Maxmin-Fair Sharing: An Example 0111131623302627 Time 29 Capacity = 6 Maxmin- fair ViFi Arrival time 0, size 72, weight 2 Arrival time 11, size 18, weight 1 Arrival time 1, size 72, weight 3Arrival time 26, size 18, weight 4
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.