HUG. Multi-Resource Fairness for Correlated and Elastic Demands. Mosharaf Chowdhury, Zhenhua Liu Ali Ghodsi, Ion Stoica

Transcription

1 HUG Multi-Resource Fairness for Correlated and Elastic Demands Mosharaf Chowdhury, Zhenhua Liu Ali Ghodsi, Ion Stoica

2 Congestion-Less Core How to share the links between multiple tenants to provide 1. optimal performance guarantees and 2. maximize utilization? L 1 L N+1 L 2 L N+2 L 3 L N+3 L N L 2N M 1 M 2 M 3 M N Tenant-A s VMs Tenant-B s VMs

3 Single-Resource Max-Min Fairness Tenant-A wants half of it Tenant-B wants all of it 1/2 1/2 L 1 1. Optimal Isolation Guarantee 1. Each tenant can create exactly one flow in each link

4 Single-Resource Max-Min Fairness Tenant-A wants half of it Tenant-B wants all of it 1/2 1/2 L 1 Progress (M) of a tenant is its demand-normalized allocation Isolation Guarantee is the minimum progress across all d A = 1/2 a A = 1/2 a A M A = = 1 d A a B d B = 1 a B = 1/2 M B = = 1/2 d B } Min(M A, M B ) = 1/2

5 Single-Resource Max-Min Fairness Tenant-A wants half of it Tenant-B wants all of it 1/2 1/2 L 1 1. Optimal Isolation Guarantee 2. Work Conservation

6 No Tradeoff for Single Resource 1. Optimal Isolation Guarantee 2. Work Conservation 3. Strategyproof Work- Conserving Utilization Per-Flow Fairness Optimal Isolation Guarantee

7 Congestion-Less Core Tenants have different 1. placements, 2. communication patterns, 3. demand correlations, 4. L 1 L N+1 L 2 L N+2 L 3 L N+3 L N L 2N M 1 M 2 M 3 M N Tenant-A s VMs Tenant-B s VMs

8 Per-Flow Fairness For Multiple Resources 1. Optimal Isolation Guarantee 2. Work Conservation 3. Strategyproof Work- Conserving Utilization Per-Flow Fairness Optimal Isolation Guarantee

9 Elastic Demands 1 Tenant-A wants all of L 1 and all of L 2 Tenant-B wants all of L 1 and all of L 2 L 1 L 2 1. FairCloud: Sharing the Network in Cloud Computing, SIGCOMM 12

10 Tenant-Level Max-Min Fairness (PS-P) 1/2 1/2 Tenant-A wants all of L 1 and all of L 2 Tenant-B wants all of L 1 and all of L 2 1/2 1/2 L 1 L 2 d A = <1, 1> d B = <1, 1> a A = <½, ½> a B = <½, ½> M A = min M B = min i a A i d A i a B ( ) ( ) i d B = ½ = ½ } Min(M A, M B ) = ½

11 Tenant-Level Max-Min Fairness (PS-P) 1. Suboptimal Isolation Guarantee 2. Work Conservation Work- Conserving Utilization Per-Flow Fairness PS-P Optimal Isolation Guarantee

12 Correlated Demands 1 Tenant-A wants some of L 1 and all of L 2 Tenant-B wants some of L 2 and all of L 1 L 1 L 2 1. Dominant Resource Fairness: Fair Allocation of Multiple Resource Types, NSDI 11

13 Dominant Resource Fairness (DRF) 1/3 2/3 Tenant-A wants exactly half unit of L 1 for each of L 2 2/3 1/9 L 1 L 2 Tenant-B wants exactly 1/6 unit of L 2 for each of L 1 d A = <1/2, 1> d B = <1, 1/6> a A = <1/3, 2/3> a B = <2/3, 1/9> i a A i d A i a B ( ) M A = min = 2/3 ( ) i M B = min = 2/3 d B } Min(M A, M B ) = 2/3

14 Dominant Resource Fairness (DRF) 1. Optimal Isolation Guarantee 2. Arbitrarily Utilization 3. Strategyproof Work- Conserving Utilization Per-Flow Fairness PS-P DRF Optimal Isolation Guarantee

15 For elastic and correlated demands, can we 1. Optimal Isolation Guarantee 2. optimal Arbitrarily isolation guarantee Utilizationand 3. Strategyproof simultaneously achieve maximum utilization? Work- Conserving Utilization PS-P Per-Flow Fairness DRF Optimal Isolation Guarantee

16 For elastic and correlated demands, can we simultaneously achieve optimal isolation guarantee and maximum utilization? Work- Conserving Utilization PS-P NO Per-Flow Fairness DRF Optimal Isolation Guarantee

17 1. Why not? For elastic and correlated 2. What s the best we demands, can we can achieve? simultaneously achieve optimal isolation 3. How can we achieve guarantee and that? maximum utilization? 4. Does it matter? Work- Conserving Utilization PS-P NO Per-Flow Fairness Optimal Isolation Guarantee DRF

18 1. Why not? For elastic and correlated 2. What s the best we demands, can we can achieve? simultaneously achieve optimal isolation 3. How can we achieve guarantee and that? maximum utilization? 4. Does it matter? Work- Conserving Utilization Per-Flow Fairness PS-P Optimal Isolation Guarantee DRF

19 Elastic and Correlated Demands 1/3 2/3 Tenant-A wants at least half unit of L 1 for each of L 2 2/3 1/9 L 1 L 2 Tenant-B wants at least 1/6 unit of L 2 for each of L 1 d A = <1/2, 1> d B = <1, 1/6> a A = <1/3, 2/3> a B = <2/3, 1/9> M A = 2/3 M B = 2/3 } Min(M A, M B ) = 2/3

20 Elastic and Correlated Demands 1/3 2/3 Who gets this? Tenant-A wants at least half unit of L 1 for each of L 2 2/3 1/9 L 1 L 2 Tenant-B wants at least 1/6 unit of L 2 for each of L 1 d A = <1/2, 1> d B = <1, 1/6> a A = <1/3, 2/3> a B = <2/3, 1/9> M A = 2/3 M B = 2/3 } Min(M A, M B ) = 2/3

21 Work Conservation Doesn t Work! 1/2 1/2 Tenant-A wants lies and at asks least half for one unit unit of Lof 1 and L 1 for each of L 2 1/2 1/12 L 1 L 2 Tenant-B wants at least 1/6 unit of L 2 for each of L 1 d' A = <1, 1> a' A = <½, ½> d B = <1, 1/6> a B = <½, 1/12>

22 Work Conservation Doesn t Work! 1/2 11/12 1/2 Tenant-A lies and asks for one unit of L 1 for each of L 2 1/2 1/12 L 1 L 2 Tenant-B wants at least 1/6 unit of L 2 for each of L 1 d' A = <1, 1> a' A = <½, ½> a' A = <½, 11/12> M' A = 11/12 d B = <1, 1/6> a B = <½, 1/12> M' B = 1/2

23 Prisoner s Dilemma Tenant-A Doesn t Lie Lies Tenant-B Doesn t Lie Lies , , ,,

24 1. Why not? For elastic and correlated Optimal isolation demands, can we guarantee depends on simultaneously achieve being strategyproof, but optimal isolation work conservation guarantee and cannot coexist with maximum utilization? strategyproof-ness Work- Conserving Utilization Per-Flow Fairness PS-P DRF Optimal Isolation Guarantee

25 1. Why not? For elastic and correlated 2. What s the best we demands, can we can achieve? simultaneously achieve optimal isolation 3. How can we achieve guarantee and that? maximum utilization? 4. Does it matter? Work- Conserving Utilization Per-Flow Fairness PS-P Optimal Isolation Guarantee DRF

26 HUG in Non-Cooperative Setting Work- Conserving PS-P 1. Optimal Isolation Guarantee 2. Highest Utilization 3. Strategyproof Utilization Per-Flow Fairness HUG DRF Optimal Isolation Guarantee

27 HUG Highest Utilization with the Optimal Isolation Guarantee Restrict a tenant s allocation in any link to its allocation in the bottleneck link

28 Tenant-A Lied 1/2 11/12 1/2 Tenant-A lies and asks for one unit of L 1 for each of L 2 1/2 1/12 L 1 L 2 Tenant-B wants at least 1/6 unit of L 2 for each of L 1 d' A = <1, 1> a' A = <½, ½> a' A = <½, 11/12> M' A = 11/12 1/2 d B = <1, 1/6> a B = <½, 1/12> M' B = 1/2

29 Everyone is Forced to Tell the Truth 1/3 2/3 Tenant-A wants at least half unit of L 1 for each of L 2 2/3 1/3 1/9 L 1 L 2 Tenant-B wants at least 1/6 unit of L 2 for each of L 1 d A = <1/2, 1> a A = <1/3, 2/3> M A = 2/3 d B = <1, 1/6> a B = <2/3, 1/9> a' B = <2/3, 1/3> M B = 2/3

30 HUG Highest Utilization with the Optimal Isolation Guarantee 1. Tenants update correlation vectors through an API 2. Operators calculate HUG centrally and enforce it locally

31 1. Why not? 2. What s the best we 1. Optimal Isolation can achieve? Guarantee 2. Highest Utilization 3. Strategyproof 3. How can we achieve that? 4. Does it matter? Work- Conserving Utilization Per-Flow Fairness PS-P Optimal Isolation Guarantee HUG DRF

32 Evaluation Per-Flow Fairness PS-P DRF HUG PS-P concurrent tenants 3000 machines with 3Tbps total capacity Original placement and communication patterns from the Facebook trace Utilization (Tbps) 1 Per-Flow Fairness HUG DRF 10-4 (Gbps) Isolation Guarantee

33 #1 #2 #3 Bursty Demands Long-Term Guarantees Decentralized Three Algorithms Challenges Periodic demand bursts in Spark streaming Predictable performance guarantees over time Survive master failures and enable low response times

34 HUG Highest Utilization with the Optimal Isolation Guarantee Generalizes single- and multi-resource fairness schemes Optimal worst-case performance guarantees for tenants Highest utilization for operators Mosharaf Chowdhury

35

36 HUG in Cooperative Setting Work- Conserving PS-P 1. Optimal Isolation Guarantee 2. Work Conservation Utilization Per-Flow Fairness HUG DRF Optimal Isolation Guarantee

37 Evaluation A 3000-machine trace-driven simulation based on a snapshot of Facebook production trace 1. Does it provide isolation guarantee? YES 2. Does it improve utilization? 3. Is it practical?

38 Optimal Progress for ALL Per-Flow Fairness PS-P 2 DRF 3 HUG Max Min Max-to-Min Progress Ratio 10000X 10X 1X 1X tenants in this particular snapshot. The unit of progress is Gbps. 2. FairCloud: Sharing the Network in Cloud Computing, SIGCOMM Dominant Resource Fairness: Fair Allocation of Multiple Resource Types, NSDI 11

39 Higher Network Utilization Total Per-Flow Fairness PS-P 2 DRF 3 HUG Utilization (Tbps) Max-to-Min Progress Ratio 10000X 10X 1X 1X tenants in this particular snapshot. The unit of progress is Gbps. 2. FairCloud: Sharing the Network in Cloud Computing, SIGCOMM Dominant Resource Fairness: Fair Allocation of Multiple Resource Types, NSDI 11

40 Long-Term Performance Fraction of Shuffles Per-Flow Fairness PS-P HUG DRF Varys Slowdown w.r.t. Minimum Required Comp. Time Average Time Per-flow Fairness 1.49X PS-P 1.14X DRF 1.14X HUG 1X Varys X 1. Efficient Coflow Scheduling with Varys, SIGCOMM 14

41 Coordination Overheads and Scalability Computation overheads Less than 5 μs for 100- machine cluster Less than 10 ms for 100,000 machines Communication overheads Less than 10 ms for 100- machine cluster Less than 1 second for 100,000 machines

42 Optimal Progress for ALL Fraction of Shuffles Per-Flow Fairness PS-P HUG DRF Progress (Gbps) Per-flow Fairness PS-P DRF HUG Max/Min Ratio 10000X 10X 1X 1X

43 Higher Utilization Fraction of Shuffles Per-Flow Fairness PS-P HUG DRF Aggregate Bandwidth (Gbps) Per-flow Fairness PS-P DRF HUG Max/Min Ratio X 2590X 196X 340X