[PPT] - Siphon: Expediting Inter-Datacenter Coflows in Wide-Area Data PowerPoint Presentation

SLIDE 1

Siphon: Expediting Inter-Datacenter Coflows in Wide-Area Data Analytics

Shuhao Liu, Li Chen, Baochun Li University of Toronto July 12, 2018

SLIDE 2

What is a Coflow?

Map 1 Map 2 Map 4 Map 3 Reduce 1 Reduce 2 Reduce 3 Reduce 4

One stage in a data analytic job

SLIDE 3

What is a Coflow?

Map 1 Map 2 Map 4 Map 3 Reduce 1 Reduce 2 Reduce 3 Reduce 4

Map tasks One stage in a data analytic job

SLIDE 4

What is a Coflow?

Map 1 Map 2 Map 4 Map 3 Reduce 1 Reduce 2 Reduce 3 Reduce 4

Map tasks Reduce tasks One stage in a data analytic job

SLIDE 5

What is a Coflow?

Map 1 Map 2 Map 4 Map 3 Reduce 1 Reduce 2 Reduce 3 Reduce 4

One stage in a data analytic job all-to-all shuffle

SLIDE 6

What is a Coflow?

Map 1 Map 2 Map 4 Map 3 Reduce 1 Reduce 2 Reduce 3 Reduce 4

One stage in a data analytic job Coflow: considered done only when all flows finish

SLIDE 7

Coflow Scheduling

Objective: minimizing average coflow completion time Network model: datacenter networking Big switch abstraction network core is congestion-free

Job 1 Job 2

SLIDE 8

Coflow Scheduling

Objective: minimizing average coflow completion time Network model: datacenter networking Big switch abstraction network core is congestion-free

Job 1 Job 2

SLIDE 9

Coflow Scheduling

Objective: minimizing average coflow completion time Network model: datacenter networking Big switch abstraction network core is congestion-free

Job 1 Job 2

SLIDE 10

Coflow Scheduling

Objective: minimizing average coflow completion time Network model: datacenter networking Big switch abstraction network core is congestion-free

Job 1 Job 2

SLIDE 11

Coflow Scheduling

Objective: minimizing average coflow completion time Network model: datacenter networking Big switch abstraction network core is congestion-free

Job 1 Job 2

SLIDE 12

Coflow Scheduling

Objective: minimizing average coflow completion time

Job 1 Job 2

1 3 2 1 2 3 1 2 3

Non-blocking Switch

Coflow 1 Coflow 2

SLIDE 13

Coflow Scheduling

Objective: minimizing average coflow completion time

Job 1 Job 2

1 3 2 1 2 3 3 2 2 3 1 1 2 3

Non-blocking Switch

Coflow 1 Coflow 2

SLIDE 14

Coflow Scheduling

Objective: minimizing average coflow completion time

Job 1 Job 2

1 3

3 2 2 3 1

2 1 2 3 3 2 2 3 1 1 2 3

Non-blocking Switch

Coflow 1 Coflow 2

SLIDE 15

Wide-Area Data Analytics

SLIDE 16

Wide-Area Data Analytics

Datacenter 1

Map 1 Map 2 Map 3 Map 4 Reduce 1 Reduce 2

Data 1

Datacenter 2

Data 2

Datacenter 3

Data 3

Datacenter 4

Data 4 Wide Area Network

SLIDE 17

Wide-Area Data Analytics

Datacenter 1

Map 1 Map 2 Map 3 Map 4 Reduce 1 Reduce 2

Data 1

Datacenter 2

Data 2

Datacenter 3

Data 3

Datacenter 4

Data 4 Wide Area Network

SLIDE 18

With tasks placed in different datacenter, what about their generated inter-datacenter coflows?

SLIDE 19

Challenges

Dumb bell network model: inter-datacenter links are the only bottleneck Datacenter A Datacenter B

Inter-datacenter link

SLIDE 20

Challenges

Constantly changing available bandwidth

Measured Bandwidth (Mbps) in a 100s interval

55 68.75 82.5 96.25 110

CA-EU US-EU

SLIDE 21

Can existing heuristics work?

Link 1 Estimated Flow Completion Time 1 2 3 4 5 6 7 8 Link 2

SLIDE 22

Coflow scheduling should consider the distribution of available bandwidth.

SLIDE 23

Monte Carlo Simulation

SLIDE 24

Monte Carlo Simulation

Scheduling Decision Tree

SLIDE 25

B [0/0]

Monte Carlo Simulation

A [0/0]

Scheduling Decision Tree

C [0/0]

SLIDE 26

B [0/0]

Monte Carlo Simulation

A [0/0]

Scheduling Decision Tree

C [0/0] B C A C B A

SLIDE 27

B [0/0]

Monte Carlo Simulation

A [0/0]

Scheduling Decision Tree

C [0/0] B C A C B A 16.2 9.3 12.5 15.5 20.2 13.1

SLIDE 28

B [0/0]

Monte Carlo Simulation

A [0/0]

Scheduling Decision Tree

C [0/0] B C A C B A 16.2 9.3 12.5 15.5 20.2 13.1

SLIDE 29

B [0/0]

Monte Carlo Simulation

A [0/0]

Scheduling Decision Tree

C [0/0] B C A C B A 16.2 9.3 12.5 15.5 20.2 13.1 B [0/1] A [1/1] C [0/1]

SLIDE 30

B [68/100]

Monte Carlo Simulation

A [29/100]

Scheduling Decision Tree

C [3/100] B C A C B A

SLIDE 31

B [68/100]

Monte Carlo Simulation

A [29/100]

Scheduling Decision Tree

C [3/100] B C A C B A

Complexity? 100 * O(n!)

SLIDE 32

Reduced Simulation Complexity

SLIDE 33

Reduced Simulation Complexity

Bounded Search Depth

Θ(t × nd)

SLIDE 34

Reduced Simulation Complexity

Bounded Search Depth Reduced Search Breath (Early termination)

Θ(t × nd)

SLIDE 35

Reduced Simulation Complexity

Bounded Search Depth Reduced Search Breath (Early termination)

Θ(t × nd) O(t × nd)

<latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit>

SLIDE 36

Reduced Simulation Complexity

Bounded Search Depth Reduced Search Breath (Early termination) Online Incremental Search

Θ(t × nd) O(t × nd)

<latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit>

SLIDE 37

Reduced Simulation Complexity

Bounded Search Depth Reduced Search Breath (Early termination) Online Incremental Search

Θ(t × nd) O(t × nd)

<latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit><latexit sha1_base64="SRO9rJSliFs1sIUx4d32DxWNFw=">AB/XicjVDLSgNBEOz1GeMrPm5eBoMQL2FXBD0GvXgzgnlAsobZ2UkyZHZ2mekV4hL8FS8eFPHqf3jzb5w8DioKFjQUVd1U0EihUHX/XDm5hcWl5ZzK/nVtfWNzcLWdt3EqWa8xmIZ62ZADZdC8RoKlLyZaE6jQPJGMDgf+41bro2I1TUOE+5HtKdEVzCKVuoUdi9JiSBpo4i4IeomC0fksFMoemV3AvI3KcIM1U7hvR3GLI24QiapMS3PTdDPqEbBJB/l26nhCWUD2uMtSxW1YX42+X5EDqwSkm6s7SgkE/XrRUYjY4ZRYDcjin3z0xuLv3mtFLunfiZUkiJXbBrUTSXBmIyrIKHQnKEcWkKZFvZXwvpU4a2sPz/SqgflT237F0dFytnszpysAf7UAIPTqACF1CFGjC4gwd4gmfn3nl0XpzX6eqcM7vZgW9w3j4BRU6T0A=</latexit>

O(t × nd−1)

<latexit sha1_base64="JK9yGuMiIKSIlpDMmqUtZVT2rQ=">AB/3icjVDLSgNBEOyNrxhfq4IXL4NBiAfDrgh6DHrxZgTzgCSG2ckGTI7u8z0CmHNwV/x4kERr/6GN/GyeOgomBQ1HVTUVxFIY9LwPJzM3v7C4lF3OrayurW+4m1tVEyWa8QqLZKTrATVcCsUrKFDyeqw5DQPJa8HgfOzXbrk2IlLXOIx5K6Q9JbqCUbRS2925JAWCpIki5Iaom7Rz6I/IQdvN+0VvAvI3ycM5b73uxELAm5QiapMQ3fi7GVUo2CST7KNRPDY8oGtMcblipq41rp5P8R2bdKh3QjbUchmahfL1IaGjMA7sZUuybn95Y/M1rJNg9baVCxQlyxaZB3UQSjMi4DNIRmjOUQ0so08L+SlifasrQVpb7XwnVo6LvFf2r43zpbFZHFnZhDwrgwmU4ALKUAEGd/AT/Ds3DuPzovzOl3NOLObfgG5+0TKpiUQg=</latexit><latexit sha1_base64="JK9yGuMiIKSIlpDMmqUtZVT2rQ=">AB/3icjVDLSgNBEOyNrxhfq4IXL4NBiAfDrgh6DHrxZgTzgCSG2ckGTI7u8z0CmHNwV/x4kERr/6GN/GyeOgomBQ1HVTUVxFIY9LwPJzM3v7C4lF3OrayurW+4m1tVEyWa8QqLZKTrATVcCsUrKFDyeqw5DQPJa8HgfOzXbrk2IlLXOIx5K6Q9JbqCUbRS2925JAWCpIki5Iaom7Rz6I/IQdvN+0VvAvI3ycM5b73uxELAm5QiapMQ3fi7GVUo2CST7KNRPDY8oGtMcblipq41rp5P8R2bdKh3QjbUchmahfL1IaGjMA7sZUuybn95Y/M1rJNg9baVCxQlyxaZB3UQSjMi4DNIRmjOUQ0so08L+SlifasrQVpb7XwnVo6LvFf2r43zpbFZHFnZhDwrgwmU4ALKUAEGd/AT/Ds3DuPzovzOl3NOLObfgG5+0TKpiUQg=</latexit><latexit sha1_base64="JK9yGuMiIKSIlpDMmqUtZVT2rQ=">AB/3icjVDLSgNBEOyNrxhfq4IXL4NBiAfDrgh6DHrxZgTzgCSG2ckGTI7u8z0CmHNwV/x4kERr/6GN/GyeOgomBQ1HVTUVxFIY9LwPJzM3v7C4lF3OrayurW+4m1tVEyWa8QqLZKTrATVcCsUrKFDyeqw5DQPJa8HgfOzXbrk2IlLXOIx5K6Q9JbqCUbRS2925JAWCpIki5Iaom7Rz6I/IQdvN+0VvAvI3ycM5b73uxELAm5QiapMQ3fi7GVUo2CST7KNRPDY8oGtMcblipq41rp5P8R2bdKh3QjbUchmahfL1IaGjMA7sZUuybn95Y/M1rJNg9baVCxQlyxaZB3UQSjMi4DNIRmjOUQ0so08L+SlifasrQVpb7XwnVo6LvFf2r43zpbFZHFnZhDwrgwmU4ALKUAEGd/AT/Ds3DuPzovzOl3NOLObfgG5+0TKpiUQg=</latexit><latexit sha1_base64="JK9yGuMiIKSIlpDMmqUtZVT2rQ=">AB/3icjVDLSgNBEOyNrxhfq4IXL4NBiAfDrgh6DHrxZgTzgCSG2ckGTI7u8z0CmHNwV/x4kERr/6GN/GyeOgomBQ1HVTUVxFIY9LwPJzM3v7C4lF3OrayurW+4m1tVEyWa8QqLZKTrATVcCsUrKFDyeqw5DQPJa8HgfOzXbrk2IlLXOIx5K6Q9JbqCUbRS2925JAWCpIki5Iaom7Rz6I/IQdvN+0VvAvI3ycM5b73uxELAm5QiapMQ3fi7GVUo2CST7KNRPDY8oGtMcblipq41rp5P8R2bdKh3QjbUchmahfL1IaGjMA7sZUuybn95Y/M1rJNg9baVCxQlyxaZB3UQSjMi4DNIRmjOUQ0so08L+SlifasrQVpb7XwnVo6LvFf2r43zpbFZHFnZhDwrgwmU4ALKUAEGd/AT/Ds3DuPzovzOl3NOLObfgG5+0TKpiUQg=</latexit>

SLIDE 38

How to enforce the scheduling decisions?

SLIDE 39

Siphon: System Overview

Form a software-defined overlay network Aggregators: measure bandwidth, schedule coflows based on priority assignment Controller: compute priority based on Monte Carlo simulation

Datacenter 1 Siphon Inter-Datacenter Software-Defined Overlay

… Worker Process Aggregator Daemon Worker Process Aggregator Daemon

Worker Process Aggregator Daemon

Datacenter 2

Worker Process Aggregator Daemon Worker Process Aggregator Daemon

Worker Process Aggregator

Daemon … Controller

SLIDE 40

Performance: Coflow Scheduling

T-V T-M V-T V-M M-T M-V 150 160 170 180 190 200 210 Available Bandwidth (Mbps) <25% 25-49% 50-74% ≥75% All Normalized CCT (%) 95.7 98.7 98.3 89.6 89.7 94.9 99.3 96.7 85.9 99.5 Average 90th Percentile

SLIDE 41

With Siphon, we can do more…

SLIDE 42

Intra-Coflow Scheduling

Largest Flow Group First

M1 M2 R2 R1 M3 50MB 100MB 50MB 50MB 50MB 50MB

DC1 to DC2

D C 3 t

D

C 2

100

DC 1-2

50

Time 1 2 3 4 Flow Group 2 ends Schedule 2

50 50 50 50

DC 3-2

200 150

R1 R2 Flow Group 1 ends 5 6 7 8

100

DC 1-2

50

Time 1 2 3 4 Flow Group 1 ends Schedule 1 (LFGFS)

50 50 50 50

DC 3-2

200 150

R1 R2 Flow Group 2 ends 5 6 7 8

SLIDE 43

Coflow Multipath Routing

35~45Mbps 8 4 ~ 8 3 M b p s 508~586Mbps

N. Carolina

5 5 5 5 5 Taiwan 5 5 5 5 5 Tokyo 5 2

3 4 4 4 4

S. Carolina

Tokyo Taiwan Oregon Belgium

216 161 216 216 198 148 197 197 149 149 149 149 227 227 227 170 206 207 155 206

SLIDE 44

Performance: Intra-Coflow Scheduling

Spark Naive Multipath Siphon 50 100 150 200 Reduce Stage Execution Time (s) 72.2 186.2 56.3 155.3 49.4 135.3 48.7 130.3 Task Execution Shuffle Read

SLIDE 45

Performance: Benchmark Workloads

ALS PCA BMM Pearson W2V FG 200 400 Application Run Time (s)

26.1
14.1
8.5
23.8
4.1

+1.7 Siphon Spark

SLIDE 46

Takeaway

Siphon is a software-defined inter-datacenter overlay that realizes: Coflow scheduling in wide-area data analytics: Monte Carlo Simulation Intra-coflow scheduling: Largest Flow Group First Coflow multipath rerouting Shorter coflow completion time leads to better job-level performance

SLIDE 47