automated)bandwidth)allocaon) problems)in)datacenters) · bandwidth)allocaon)problem ) 1gbps 600m...
TRANSCRIPT
Automated Bandwidth Alloca1on Problems in Data Centers
Yifei Yuan, Anduo Wang, Rajeev Alur, Boon Thau Loo University of Pennsylvania
Mo1va1on
• Managing network resources is the key computa1onal problem in Data Centers.
• Applying verifica1on/synthesis tool to network resource management? – Benefits: exact solu1ons, correctness guarantees – Challenges: efficiency
• This work: bandwidth alloca1on by SAT/SMT solvers
Bandwidth Alloca1on Problem
Bandwidth Alloca1on Problem
1G bps 600M bps
500M bps
450M bps
X1
X2 X3
S1 S2 S3 S4
Data Center’s Network
10G bps
10G bps
Bandwidth Alloca1on Problem
1G bps 600M bps
500M bps
450M bps
X1
X2 X3
S1 S2 S3 S4
Data Center’s Network
V1
V2 V3
400M bps
400M bps
Virtual Network
10G bps
10G bps
Bandwidth Alloca1on Problem
1G bps 600M bps
500M bps
450M bps
X1
X2 X3
S1 S2 S3 S4
Data Center’s Network
V1
V2 V3
400M bps
400M bps
Virtual Network
10G bps
10G bps
Bandwidth Alloca1on Problem
1G bps 600M bps
500M bps
450M bps
X1
X2 X3
S1 S2 S3 S4
Data Center’s Network
V1
V2 V3
400M bps
400M bps
Virtual Network
v1 v3 v2
10G bps
10G bps
Bandwidth Alloca1on Problem
1G bps 600M bps
500M bps
450M bps
X1
X2 X3
S1 S2 S3 S4
Data Center’s Network
V1
V2 V3
400M bps
400M bps
Virtual Network
v1 v3 v2
10G bps
10G bps
BAP: Facts
• Complexity: – NP-‐complete: tree for physical network & virtual network
• Exis1ng heuris1cs: – Pros: efficient
– Cons: no guarantee • Alterna1ve approach: SAT/SMT solving
SAT Encoding: A Glimpse
• X(v,s): VM v is mapped to server s • Y(l,e): physical link l is reserved bandwidth virtual link e
• R(l,e,k): physical link l is the k-‐th edge on the rou1ng path for virtual link e
• Server capacity: – ∑v X(v,s) < c(s), for every server s
• Link capacity: – ∑e Y(l,e) < b(l), for every physical link l
Abstrac1on and Refinement
• Observa1on: Hierarchical physical network topology in data centers – Tree – Fat-‐tree
• Idea: – Abstract physical network: small size – Refine subgraphs
Abstrac1on
1 2 4 2
Abstrac1on
1 2 4 2
Abstrac1on
1 2 4 2
3 6
Abstrac1on
3 6
Abstrac1on
3 6
Refinement
1 2 4 2
3 6
Refinement
1 2 4 2
Evalua1on: Set up
• Physical network topology: tree with 200 servers:
200
20
4 4 4 4
Evalua1on: Set up
• Virtual network topology: connected cliques
2
1 1
Evalua1on: Set up
• Experiment: – Run alloca1on algorithm
– Keep mapping the VN to the PN – Stop when no more VN can be mapped
Evalua1on: Server U1liza1on
0
0.2
0.4
0.6
0.8
1
1.2
9 vms 15 vms
Avg. server
uli,za,on
# of VMs
secondnet sat sat_abs
Evalua1on: Link U1liza1on
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
9 vms 15 vms Avg. link u,
liza,
on
# of VMs
secondnet sat sat_abs
Evalua1on: Running Time per VN
0.01
0.1
1
10
100
9 vms 15 vms Runn
ing ,me pe
r vn (secon
ds)
# of VMs in the virtual network
secondnet sat sat_abs
Summary
• Alterna1ve approach solving network resource alloca1on problem: using SAT/SMT solvers
• Abstract&refinement for scalability
• Strength: op1mal solu1on
• Weakness: efficiency – Possible scenario: Op1mal realloca1on