1. Phurti: Application and Network-Aware Flow Scheduling for Multi-Tenant MapReduce Clusters
Chris Cai, Shayan Saeed, Indranil Gupta, Roy Campbell, Franck Le
Systems Research Group
Distributed Protocols Research Group

2. Outline Introduction System Architecture Scheduling Algorithm
Evaluation
Summary

3. Multi-tenancy in MapReduce Clusters
Jobs
Users
MapReduce Cluster
Better ROI, high utilization.
How to share resources?
Network is the primary bottleneck.

4. Problem Statement
How to schedule network traffic to improve completion time for MapReduce jobs?

5. Application-Awareness in Scheduling
Job 1 Traffic
Job 2 Traffic
Link 1
6 units
Link 2
3 units
2 units
Fair Sharing1
Shortest Flow First2
Application Aware L1 L1 L1 L2 L2 L2 2 4 6
time 2 4 6
time
time 2 4 6
Job 1 Completion time = 5
Job 1 Completion time = 5
Job 1 Completion time = 3
Job 2 Completion time = 6
Job 2 Completion time = 6
Job 2 Completion time = 6 5

6. Network-Awareness in Scheduling
Path 1 N1 S1 S2 N3 N2 N4
Path 2
Job 1 Traffic
Job 2 Traffic
Path 1
3 units
Path 2
3 units

7. Network-Awareness in Scheduling
Job 1 Traffic
Job 2 Traffic
Path 1
3 units
Path 2
3 units
Network-Agnostic
Network-Aware P1 P1 P2 P2 2 4 6
time 2 4 6
time
Job 1 Completion time = 6
Job 1 Completion time = 3
Job 2 Completion time = 6
Job 2 Completion time = 6
Takeaway: Do not schedule interfering flows of concurrent jobs together

8. Related Work Traditional flow-scheduling
PDQ [SIGCOMM ‘12], Hedera [NSDI ‘10]
Only improve network-level metrics
Application and Network-Aware Task Schedulers
Cross-Layer Scheduling [IC2E 2015], Tetris [SIGCOMM ’14]
Schedule tasks instead of network traffic
Application-Aware traffic schedulers
Baraat [SIGCOMM ‘14], Varys [SIGCOMM ’14]
Unaware of network topology

9. Phurti: Contributions
Improves Job Completion Time
Fairness and Starvation Protection
Scalable
API Compatibility
Hardware Compatibility

10. Outline Introduction System Architecture Scheduling Algorithm
Evaluation
Summary

11. Phurti Framework Hadoop Nodes N4 N5 N6 N1 N2 N3 Phurti
Northbound API N4 N5 N6 N1 N2 N3
Phurti
Scheduling Framework
Southbound API
SDN Switches S1 S2

12. Outline Introduction System Architecture Scheduling Algorithm
Evaluation
Summary

13. Phurti Algorithm – Intuition
Job 1 Flows
Job 2 Flows 1 3 1 3 2 4 2 4
Max. Sequential Traffic: 4 units
Max. Sequential Traffic: 5 units P1 P1 P2 P2 2 4 6
time 4
time 2 6
Job 1 Completion time = 4
Job 2 Completion time = 5
Takeaway: Job completion time is determined by maximum
sequential traffic.

14. Phurti Algorithm – Intuition (cont.)
Job 1 Traffic 1 3
Max. Sequential Traffic: 4 units
Job 2 Traffic
Max. Sequential Traffic: 5 units 2 4
If Job 2 scheduled first
If Job 1 scheduled first P1 P1 P2 P2 4
time 2 6 8 2 4 6 8
time
Job 1 Completion time = 4
Job 1 Completion time = 8
Job 2 Completion time = 8
Job 2 Completion time = 5
Observation: It is better to schedule the job with smaller maximum
sequential traffic first.

15. Phurti Algorithm
Assign priorities to jobs based on Max Sequential Traffic N2 N3 s1 s2 s3 N4 N1
Let flows of the highest priority job transfer N1 N4 N1 N2 N3 N4
Let non-interfering flows of the lower priority jobs transfer
Job
Flow
Size
Max Seq. Traffic
Priority J1
N1N4 2
LOW
N4N1 J2
N2N3 1
HIGH
Let other lower priority flows transfer at a small rate
Latency Improvement
Throughput Maximization
Starvation Protection

16. Evaluation Baseline: Fair Sharing (Default in MapReduce)
Testbed: 6 nodes, 2 SDN switches
SWIM workload: workload generated from Facebook Hadoop trace
Job Size Bin
% of total jobs
% of total bytes in shuffled data
Small
62%
5.5%
Medium
16%
10.3%
Large
22%
84.2%

17. Job Completion Time
95% of jobs have better job completion time under Phurti.
Negative values mean Phurti performs better.

18. Job Completion Time
13% improvement in 95th percentile job completion time showing starvation protection.
Much better for smaller jobs since they typically have higher priority

19. Flow Scheduling Overhead
Simulate a fat-tree topology with 128 hosts.
Even in unlikely event of 100 simultaneous incoming flows, scheduling time is 4.5ms which is negligible scheduling overhead.

20. Flow Scheduling Overhead
Scheduling time for a new flow with 10 ongoing flows in the network
Scheduling overhead grows much slower than linear rate showing that it is scalable with increasing number of hosts.

21. Phurti vs Varys
Simulate 128-hosts fat-tree topology with core network having 1x, 5x and 10x capacity compared to access links
Outperforms Varys significantly when the core network has much less capacity (oversubscribed).
Better than Varys in every case.

22. Phurti: Contributions
Improves completion time for 95% of the jobs, decreases the average completion time by 20% for all jobs.
Fairness and Starvation Protection. Improves tail job completion time by 13%.
Scalable. Shown to scale to 1024 hosts and 100 simultaneous flow arrivals.
API Compatibility
Hardware Compatibility

23. Backup slides

24. Effective Transmit Rate
80% of jobs have effective transmit rate larger than 0.9 showing minimal throttling.