Decentralised Diffusion-based Quota Management Rüdiger Kapitza, Meik Felser , Hans P. Reiser, Franz J. Hauck AspectIX Research Group University Erlangen-Nuremberg Motivation � Grid infrastructures provide resources � Resources are shared among multiple parties � Resource access has to be restricted � Otherwise a user could allocate all resources • Accident • Attack • Unawareness � Solution: Quotas Decentralised Diffusion-based Quota Management 2 1
Motivation � Existing implementation of quota-based approaches 1. Quota manager issues fractions of quota to the clients 2. Client can acquire resources up to the quota � Central quota manager is the bottleneck Large-scale grids • Frequent quota allocations • Single point of failure • Decentralised Diffusion-based Quota Management 3 Motivation � decentralised quota management system � No central quota service • Scalability � Distributed quota information • Fault-tolerance Decentralised Diffusion-based Quota Management 4 2
Outline � Motivation � Diffusive Balancing • Load Balancing • Quota Balancing • Basic Protocol � Extension for Fault-Tolerance � Handling of Non-Refundable Quotas � Conclusion and Ongoing Work Decentralised Diffusion-based Quota Management 5 Diffusive Load Balancing � Load balancing for distributed multiprocessor systems (Cybenko, 1998) � Nodes are organised in small overlapping groups � Load information is exchanged � If there is an imbalance, load is migrated • A , B represent nodes • w A , w B denote the current load of nodes A and B • If w A < w B then move load ( w A - w B ) / 2 from B to A and vice versa if w A > w B A B Decentralised Diffusion-based Quota Management 6 3
Diffusive Quota Balancing: Basic Idea � Quota management: • Do NOT balance load • Balance the free quota information � Quota information is distributed � Fast resource allocation � Resource demand may be fulfilled locally Decentralised Diffusion-based Quota Management 7 Diffusive Quota Balancing: Basic Idea � Free quota is distributed and balanced among all nodes • Initial distribution of free quota Node A Node B Node C free 60 60 quota Decentralised Diffusion-based Quota Management 8 4
Diffusive Quota Balancing: Basic Idea � Free quota is distributed and balanced among all nodes • Application has been started • Resource demand was fulfilled immediately Node A Node B Node C 30 60 60 used Decentralised Diffusion-based Quota Management 9 Diffusive Quota Balancing: Basic Idea � Free quota is distributed and balanced among all nodes • System reached quota equilibrium again • Free quota is balanced again Node A Node B Node C 50 50 50 Decentralised Diffusion-based Quota Management 10 5
Basic Quota Balancing Protocol � Quota Manager: • Initial node • Injection point of initial quota • Not a central component quota manager Decentralised Diffusion-based Quota Management 11 Basic Quota Balancing Protocol � Quota manager � Initial set: • Connected to the quota manager • Connected to some other nodes of the initial set • Forming an unstructured network quota manager initial set Decentralised Diffusion-based Quota Management 12 6
Basic Quota Balancing Protocol � Quota manager � Initial set � Additional nodes: connect to a random subset of the nodes • quota manager initial set new nodes Decentralised Diffusion-based Quota Management 13 Basic Quota Balancing Protocol � Each node stores its local free quota 0 0 0 0 0 0 0 0 0 0 0 0 Decentralised Diffusion-based Quota Management 14 7
Basic Quota Balancing Protocol � Each node stores its local free quota � Global quota is injected into the quota manager 300 300 300 0 0 0 0 0 0 0 0 0 0 Decentralised Diffusion-based Quota Management 15 Basic Quota Balancing Protocol � If the local free quota changes • The nodes exchange quota information and • Balance quota differences 150 150 150 0 0 150 150 0 0 0 0 0 0 Decentralised Diffusion-based Quota Management 16 8
Basic Quota Balancing Protocol � If the local free quota changes • The nodes exchange quota information and • Balance quota differences � This continues … 75 75 75 75 75 75 75 0 0 0 0 75 75 75 Decentralised Diffusion-based Quota Management 17 Basic Quota Balancing Protocol � If the local free quota changes • The nodes exchange quota information and • Balance quota differences � This continues until the equilibrium is reached 50 50 50 50 50 50 50 50 50 50 50 50 Decentralised Diffusion-based Quota Management 18 9
Basic Quota Balancing Protocol � Resource demands reduce the free local quota • This can result in negative free local quota � The algorithm starts balancing the free quota 50 50 50 demands 60 50 -10 -10 50 50 50 50 50 50 Decentralised Diffusion-based Quota Management 19 Basic Quota Balancing Protocol � The algorithm is balancing the free quota 20 20 30 50 50 20 20 50 50 50 50 50 50 Decentralised Diffusion-based Quota Management 20 10
Basic Quota Balancing Protocol � Eventually … 35 35 15 50 50 35 35 35 35 50 50 15 35 35 Decentralised Diffusion-based Quota Management 21 Basic Quota Balancing Protocol � Eventually the system is in equilibrium again 40 40 40 40 40 40 40 40 40 40 40 40 Decentralised Diffusion-based Quota Management 22 11
Basic Quota Balancing Protocol � Node failures lead to the loss of quota information 50 50 50 50 node failure 50 50 50 50 50 50 50 50 50 50 50 50 ∑ : 200 ∑ : 150 Decentralised Diffusion-based Quota Management 23 Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter 300 0 300 300 0 0 0 0 0 0 0 0 0 0 0 0 0 Decentralised Diffusion-based Quota Management 24 12
Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter • Outgoing quota units are added to the link counter • Incoming quota units are subtracted from the link counter +150 150 150 -150 0 150 150 150 0 0 0 0 0 0 0 0 0 Decentralised Diffusion-based Quota Management 25 Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter • Positive values represent outgoing quota units • Negative values represent incoming quota units +150 150 150 -150 0 75 75 +75 0 -75 75 75 75 0 0 0 0 Decentralised Diffusion-based Quota Management 26 13
Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter • Positive values represent outgoing quota units • Negative values represent incoming quota units +150 112 112 -150 +38 38 75 75 +75 -38 -75 113 113 0 0 0 0 Decentralised Diffusion-based Quota Management 27 Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter • Positive values represent outgoing quota units • Negative values represent incoming quota units +169 93 93 -169 +38 19 94 94 +75 -38 -75 56 56 57 +57 57 57 -57 Decentralised Diffusion-based Quota Management 28 14
Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter � Node failure +169 93 93 -169 +38 94 94 +75 -38 -75 56 56 +57 57 57 -57 Decentralised Diffusion-based Quota Management 29 Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter used for quota reconstruction! • On node failure link counter is added to local quota. +169 93 93 +38 -38 -75 56 56 +57 57 57 -57 Decentralised Diffusion-based Quota Management 30 15
Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter used for quota reconstruction! • On node failure link counter is added to local quota. • This may result in negative local quota 262 262 +38 -38 -19 -19 +57 57 57 -57 Decentralised Diffusion-based Quota Management 31 Extension for Fault-Tolerance � Spontaneous node failures should be tolerated • Without losing quota information � Solution: link counter used for quota reconstruction! • On node failure link counter is added to local quota. • This may result in negative local quota 121 121 +179 141 -179 122 122 +57 57 57 -57 Decentralised Diffusion-based Quota Management 32 16
Failure of the Quota Manager � Quota manager fails • Initial node set does not evaluate link counter • Quota manager is restored via passive replication � System can run without the quota manager +169 93 93 -169 +38 94 94 +75 -38 -75 56 56 +57 57 57 -57 Decentralised Diffusion-based Quota Management 33 Failure of the Quota Manager � Quota manager fails • Initial node set does not evaluate link counter • Quota manager is restored via passive replication � System can run without the quota manager -169 94 94 +75 -38 -75 56 56 +57 57 57 -57 Decentralised Diffusion-based Quota Management 34 17
Recommend
More recommend