SLIDE 1
Networked Caches in Real-World Web Stack
Web Servers Backend Store Caches
Get X Get X X X
Characterizing Load Imbalance in Real-World Networked Caches Qi - - PowerPoint PPT Presentation
Characterizing Load Imbalance in Real-World Networked Caches Qi - - PowerPoint PPT Presentation
Characterizing Load Imbalance in Real-World Networked Caches Qi Huang Cornell U, Facebook Helga Gudmundsdottir Emory U, Reykjavik U Ymir Vigfusson Emory U, Reykjavik U Daniel A. Freedman Technion Ken Birman Cornell U Robbert van Renesse
SLIDE 2
SLIDE 3
Partitioning Data
Hash(A)
id: A type: USER name: Helga
SLIDE 4
Partitioning Data
Hash(1)
Shard 1
A
SLIDE 5
- Hashing schemes balance quantity of content per server.
- However, content popularity varies!
- Need to provision servers to handle peak load.
- Lack of real-world data to verify suspicions.
Causes of Load Imbalance
SLIDE 6
Questions
- What is the state of load imbalance?
- What contributes to load imbalance?
- How effective are current techniques?
- How might they be improved?
Sampled production traffic in TAO, serving Facebook’s social graph.
SLIDE 7
What is the State of Load Imbalance?
Significant load imbalance observed in TAO.
SLIDE 8
Skewed content popularity observed at both shard and object level.
Possible Causes: Content Popularity
SLIDE 9
Possible Causes: Hot Objects
Very hot objects alone are not a major cause of load imbalance.
SLIDE 10
Popular shards receive significantly higher load, compared to hot objects.
Possible Causes: Hot Shards
SLIDE 11
How Effective are Current Techniques?
- Hashing: Balance number of shards across servers.
- Replication: Divide load of popular shards among servers.
- Metric: Maximum load versus average.
Replication Hashing libketama TAO Perfect None TAO Perfect
SLIDE 12
- Where We Are Today
Replication Hashing libketama TAO Perfect None 1.53 TAO Perfect
SLIDE 13
- Where We Are Today
Replication Hashing libketama TAO Perfect None 1.53 TAO Perfect 1.00
SLIDE 14
- Where We Are Today
Replication Hashing libketama TAO Perfect None 1.53 TAO 1.25 Perfect 1.00
SLIDE 15
- Hashing Schemes w/o Replication
Replication Hashing libketama TAO Perfect None 1.53 TAO 1.25 Perfect 1.00
SLIDE 16
- Hashing Schemes w/o Replication
Replication Hashing libketama TAO Perfect None 1.53 1.46 TAO 1.25 Perfect 1.00
SLIDE 17
- Hashing Schemes w/o Replication
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.25 Perfect 1.00
SLIDE 18
- libketama Hashing
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.25 Perfect 1.00
SLIDE 19
- libketama Hashing
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.53 1.25 Perfect 1.00
SLIDE 20
- libketama Hashing
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.53 1.25 Perfect 1.41 1.00
SLIDE 21
- TAO: Room For Improvement
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.53 1.25 Perfect 1.41 1.00
SLIDE 22
- TAO: Room For Improvement
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.53 1.25 Perfect 1.41 1.18 1.00
SLIDE 23
- TAO: Room For Improvement
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.53 1.25 1.17 Perfect 1.41 1.18 1.00
SLIDE 24
- Streaming Replication
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.53 1.25 1.17 Perfect 1.41 1.18 1.00
SLIDE 25
- Streaming Replication
Replication Hashing libketama TAO Perfect None 1.53 1.46 1.34 TAO 1.53 1.25 1.17 Perfect 1.41 1.18 1.00 Streaming with Perfect Hashing achieves max/avg 1.12
SLIDE 26
Summary and Future Work
- Characterized how hashing and replication
affect load imbalance.
- Can streaming algorithms replicate content
before its popularity surges?
- Can we predict popularity spikes and