Download presentation
Presentation is loading. Please wait.
Published byEmmeline Collins Modified over 8 years ago
1
SmartGRID Decentralized, dynamic grid scheduling framework on swarm agent-based intelligence GCC'08, shenzhen, China. Oct. 26, 2008 Ye HUANG, Amos BROCCO Grid Group, Dept of Information and Communication Technologies, EIA-FR, Switzerland Pervasive Artificial Intelligence Group, Dept of Informatics, University of Fribourg, Switzerland
2
2 Outline Aim SmartGRID architecture SmartGRID in depth MaGate scheduler Data warehouse interface Agent-based swarm intelligence Summary Future work
3
3 Aim Vision of Grid: Large scale distributed resources Decentralized Different policies Unstable, low reliability SmartGRID: grid resource management framework Utilizing different scheduling algorithms Interoperation emphasized scheduler Dynamic resource discovery by swarm intelligent algorithm
4
4 SmartGRID layered architecture Loosely coupled layered architecture. Two layers and one internal interface. Smart Resource Management Layer Data Warehouse Interface Smart Signaling Layer
5
5 SmartGRID Node (SG-Node) SG-Node: the logical unit of SmartGRID framework MaGate scheduler, DW interface, Ant nest MaGate Scheduler Nest Info. collector DW Interface
6
6 MaGate scheduler MaGate stands for “Magnetic Gateway for scheduler” Core of Smart Resource Management Layer (SRML) Targets: Open platform to different scheduling algorithms Based on dynamic infrastructure information Gateway between schedulers (MaGate & others, etc. PBS, MSS) Allow heterogenous and dynamic Grid scheduling Manage community of resources Decentralized view of the Grid Interface to external services To deal with corresponding issues, e.g. network behavior analyzing MaGate focuses on: Decentralization & Interoperability
7
7 MaGate architecture 1. Self-management 2. Access to invoker 3. Community management 4. LRM utilization 5. External components Remote MaGate Local Resource Management Grid applications
8
8 MaGate behavior MaGate Community Job executor Interface to invoker Router Interface to external service Full functional package
9
9 MaGate current roadmap Community Component Interoperation protocol & behavior, negotiation model DRM Component SAGA API Interface Component Application-Interface (App-I) to POP-C++ External Component Multi- scheduling algorithm adoption mechanism Utilizing agent-based dynamic resource discovery
10
10 MaGate technical timetable for current roadmap Simulation Basic on GridSim, integrated with Alea, GSSIM (Done) DataWarehouse Interface prototype Communication channel between SRML and SSL (Ongoing) Candidate standard specification Scheduler Interoperability best practice, WS-Agreement, JSDL, CSG (Ongoing) First standalone prototype
11
11 Smart Signaling Layer What do we need? A network and remote resources information source » network monitoring » resource discovery Other Requirements – adaptive, reliable, robust behavior over unstable network
12
12 Smart Signaling Layer What do we need? A network and remote resources information source » network monitoring » resource discovery Other Requirements – adaptive, reliable, robust behavior over unstable network Swarm Intelligence Algorithms
13
13 Ant-based swarm intelligence Swarm intelligence? Artificial intelligence inspired by the behavior of swarms (of insects) » Typically used for optimization problems (Particle Swarm Optimization, Ant Colony Optimization) ... »...but also suitable for fully distributed algorithms Our approach Use ants (lightweight mobile agents traveling across the network) to perform different tasks on the network.
14
14 Why swarm intelligence? Strenghts? – Fully distributed algorithm, asynchronous communication Ants can only access local resources on nodes Collaboration between individuals through indirect communication (stigmergy, pheromone trails) – Robust and fault tolerant algorithms loss of individuals can be tolerated – Adaptive behavior adapts to changing network environment
15
15 Datawarehouse interface – Why use a DWI? Loosely coupled communication between layers – Easier to adapt the framework to different scenarios – In detail... Persistent and cached Grid information storage – local information monitored by the scheduler – coordinated scheduling information negotiated by scheduler – network information gathered by ants – service request queries
16
16 Goals Mission: – Membership Management and Resource Discovery
17
17 Goals Mission: – Membership Management and Resource Discovery Idea: – Maintain an optimized topology across nodes: – lower TTL for resource discovery queries, lower communication overhead
18
18 Goals Mission: – Membership Management and Resource Discovery Idea: – Maintain an optimized topology across nodes: – lower TTL for resource discovery queries, lower communication overhead Implementation: – BlåtAnt algorithm
19
19 BlåtAnt Algorithm: introduction BlåtAnt – Fully distributed algorithm using ant colonies to construct and maintain a peer-to-peer overlay topology with bounded diameter – Pure peer-to-peer, unstructured networks: » simple membership management – Balanced link distributions: no large hubs – Fault resilient
20
20 BlåtAnt Algorithm: logic Goal – Ensure that the network diameter d is D ≤ d < 2D – 1 How? – Create and remove logical links: Connection Rule: two nodes are connected if their distance (number of hops) is greater than 2D – 1 Disconnection Rule: two adjacent nodes are disconnected if there exist an alternative path between them of length less than D
21
21 BlåtAnt Algorithm: connection rule (1) – Connection Rule: two nodes are connected if their distance (number of hops) is greater than 2D – 1 A B D = 4 2D – 1 = 7 d(A,B) = 8 > 2D - 1 = 7
22
22 BlåtAnt Algorithm: connection rule (2) – Connection Rule: two nodes are connected if their distance (number of hops) is greater than 2D – 1 A B D = 4 2D – 1 = 7
23
23 BlåtAnt Algorithm: disconnection rule (1) – Disconnection Rule: two adjacent nodes are disconnected if there exist an alternative path between them of length less than D A C B D = 4 2D – 1 = 7 Alternative path d(A,C) = 3 < D = 4
24
24 BlåtAnt Algorithm: disconnection rule (2) – Disconnection Rule: two adjacent nodes are disconnected if there exist an alternative path between them of length less than D A C B D = 4 2D – 1 = 7 Not necessary!
25
25 BlåtAnt Algorithm: disconnection rule (3) – Disconnection Rule: two adjacent nodes are disconnected if there exist an alternative path between them of length less than D A C B D = 4 2D – 1 = 7
26
26 BlåtAnt Algorithm: implementation Who does what? – Different species of ants with different tasks: Collect and spread information Connect / Disconnect peers
27
27 BlåtAnt Algorithm: implementation Who does what? – Different species of ants with different tasks: Collect and spread information Connect / Disconnect peers – Nodes run computation tasks: Discover nodes matching connection/disconnection rules based on a partial view of the network
28
28 BlåtAnt Algorithm: an example graph before augmentation graph after augmentation diameter = 4 (< 2D – 1, D=3) diameter = 19
29
29 Other details Where do ants live? – Solenopsis Framework: fully distributed platform to execute ant algorithms Sandboxed environment Extensible Support for strong migration of ant agents
30
30 Summary Smart Resource Management Layer – modular scheduler architecture – decentralized, distributed – open to existing local schedulers and external services Data Warehouse Interface Smart Signaling Layer – ant algorithms to provide network services – underlying peer-to-peer topology constructed using the BlåtAnt algorithm
31
31 Future work MaGate Scheduler – First prototype – SG-Node validation Data Warehouse – First development and integration Smart Signaling Layer (SSL) – Further research on resource discovery algorithms, round-trip time optimization, pro- active monitoring – Solenopsis 2.0 Middleware
32
32 Thanks! Questions?
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.