Download presentation
Presentation is loading. Please wait.
Published byGregory Dylan Murphy Modified over 9 years ago
1
Distributed Load Balancing for Key-Value Storage Systems Imranul Hoque Michael Spreitzer Malgorzata Steinder
2
2
3
Key-Value Storage Systems Usage: – Session state, tags, comments, etc. Requirements: – Scalability – Fast response time – High availability & fault tolerance – Relaxed consistency guarantee Example: Cassandra, Dynamo, PNUTS, etc. 3
4
Load Balancing in K-V Storage Hash partitioned vs. range partitioned – Range partitioned data ensures efficient range scan/search – Hash partitioned data helps even distribution Server 1 Server 2 Server 3 Server 4 SAT TUE SUN MON WED THU FRI MON TUE WED THU FRI SAT SUN Tablets Table 4
5
Issues with Load Balancing Uneven space distribution due to range partitioning – Solution: partition the tablets and move them around Few number of very popular records Server 1 Server 2 Server 3 Server 4 SAT TUE SUN MON WED THU FRI 5
6
Contribution Algorithms for solving the load balancing problem – Load = space, bandwidth – Evenly distribute the spare capacity – Distributed algorithm, not a centralized one – Reduce the number of moves Previous solutions: – One dimensional/key-space redistribution/bulk loading 6
7
Outline Motivation System modeling and assumptions Algorithms – One-to-one – One-to-n – Move suppression Design decisions Experimental results Emulation of proposed distributed algorithms Future works 7
8
System Modeling and Assumptions Table Tablet Server A Server B Server C B 1, S 1 B 2, S 2 B 3, S 3 B A, S A B B, S B B C, S C 8 1.<= 0.01 in both dimensions 2. # of tablets >> # of nodes 1.<= 0.01 in both dimensions 2. # of tablets >> # of nodes B 1, S 1 B 4, S 4 B 5, S 5
9
System State B B SS Target Zone: helps achieve convergence Target Point Goal: Move tablets around so that every server is within the target zone 9
10
Load Balancing Algorithms Phase 1: – Global averaging scheme – Variance of the approximation of the average decreases exponentially fast Phase 2: – One-to-one gossip – One-to-n gossip – Move suppression Phase 1 Phase 2 Phase 1 Phase 2 t 10
11
One-to-One Gossip Point selection strategy – Midpoint strategy – Greedy strategy Tablet transfer strategy – Move to the selected point with minimum cost (space transferred) 11
12
Tablet Transfer Strategy Server 2 Server 1 Target for Server 1 B B SS 12
13
Tablet Transfer Strategy (2) Server 1 Left Right Start with an empty bag Goal: take vectors from the servers so that they add up to the target vector If slope(bag + left + right) < slope(target): – Add right to bag, move right – Otherwise, add left to bag move left 13
14
Initial Configurations Uniform Two Extreme Mid Quadrant 14
15
Point Selection Strategy Midpoint Strategy + Guaranteed convergence + No need to run phase 1 – Lots of extra movement Visualization Demo – Uniform Uniform – Two extreme Two extreme – Mid quadrant Mid quadrant SS B B Server 1 Server 2 15
16
Point Selection Strategy (2) Greedy Strategy – Take the point closer to the target – Move it to the target, if improves the position of the other point does not worsen by more than δ Reduces movement Server 1 Server 2 Takes long time to converge in some casessome Takes long time to converge in some casessome 16
17
DHT-based Location Directory 17
18
DHT + Midpoint Greedy + fallback to DHT: – Convergence problem exists for some configurations – Visualization Demo Visualization Demo Solution: – Greedy + fallback to DHT with Midpoint – Demo: uniform, two extreme, mid quadrantuniformtwo extrememid quadrant Alternate approach: – Greedy + fallback to Midpoint – Trade-off: movement cost vs. DHT overhead 18
19
Experimental Evaluation Uniform configuration – Greedy + DHT (Midpoint) – Midpoint – Greedy + Midpoint (No DHT) Effect of varying target zone Effect of failed gossip count Metrics – Amount of space moved – # of gossip rounds – Multiple tablet move 19
20
Uniform Configuration: Results 20
21
Effect of Varying Target Zone Larger target zone = fast convergence, less accuracy 21 Target zone width should depend on the target point value
22
Effect of Failed Gossip Count (Greedy) Large failed gossip count = More time in greedy mode, more unproductive gossip at the end 22
23
One-to-N Gossip Contact a few random nodes – Locked/unlocked mode Pick the most profitable one – Distance from the target is minimized Advantage – Better choices Initial results – Locked mode: may lead to deadlock – Unlocked mode: most of the cases other nodes start transfer 23
24
Move Suppression Two global stages Stage 1: – One-to-One gossip, but moves are hypothetical Stage 2: – Change to chosen placement Advantage – Tablet not moved multiple times Challenges – When to switch to Stage 2 from Stage 1 24
25
Future Works Handling initial placement Frequency of running the placement algorithm Considering the network hierarchy Handling failures Extending to heterogeneous resources Questions? 25
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.