Divya Gupta*, Lucas Perronne*, Sara Bouchenak+

*Univ. Grenoble Alpes, LIG Grenoble, France, +Univ. Lyon, INSA Lyon, LIRIS, Lyon, France

divya.gupta@imag.fr, lucas.perronne@imag.fr, sara.bouchenak@insa-lyon.fr

BFT-Bench: A Framework to Evaluate BFT ProtocolsLucas Perronne

7th ACM/SPEC International Conference on Performance Engineering – 14/03/2016

Cloud Computing Outages

2 BFT-Bench: A Framework to Evaluate BFT Protocols

Fault Tolerance through State Machine Replication

request

reply

� The same state must be shared by correct replicas� Safety

� Correct requests must be eventually executed� Liveness

3 BFT-Bench: A Framework to Evaluate BFT Protocols

Byzantine Fault Tolerance

Malicious behaviors

4

Network failureNode crash

� Providing Byzantine Fault Tolerance is the ability to ensure Liveness and Safety in presence of any byzantine components.

BFT-Bench: A Framework to Evaluate BFT Protocols

Problem Statement

� Byzantine Fault Tolerance - Related works� BFT From Theory to Practice

� PBFT

� Performance Improvement in Fault-Free Conditions� Zyzzyva, Aliph, …

� Performance Improvement in Presence of Faults� Aardvark, RBFT, …

� Evaluating BFT protocols� Usage of micro-benchmarks

� Ad-hoc settings for evaluations

� No benchmarking environment dedicated to BFT

5 BFT-Bench: A Framework to Evaluate BFT Protocols

BFT-Bench

1. Several protocols implementations

2. Various fault scenarios

6

3. Configurable workload

4. Statistics monitoring

BFT-Bench: A Framework to Evaluate BFT Protocols

Network flooding

WorkloadNumber of clients: 10Request size: 4KbComputational time: 30 ms

FaultloadType of fault: Network FloodingFault trigger time: 300sFaulty node: Non-primary replicaMessage size: 4Kb

7 BFT-Bench: A Framework to Evaluate BFT Protocols

Replica Crash

WorkloadNumber of clients: 5Request size: 4KbComputational time: 10 ms

FaultloadType of fault: Replica CrashFault trigger time: 300sFaulty node: Primary replica

8 BFT-Bench: A Framework to Evaluate BFT Protocols

Message delay

WorkloadNumber of clients: 2Request size: 4KbComputational time: 100 μs

FaultloadType of fault: Message delayFault trigger time: 300sFaulty node: Primary replicaDelay time: 100 ms

9 BFT-Bench: A Framework to Evaluate BFT Protocols

Conclusion & Perspectives

� BFT-bench includes several state-of-the-art BFT protocols

� It implements many faulty behaviors on different workloads

Perspectives

• Additional Protocols

10

Additional Protocols

• More faulty scenarios

• https://bftbench.gforge.liris.cnrs.fr/

BFT-Bench: A Framework to Evaluate BFT Protocols