my point of view about bandwidth sharing
DESCRIPTION
My Point of View about Bandwidth Sharing. Bin Wang. Outline. Oktopus (Sigcomm 2011) TIVC (Sigcomm 2012) Seawall (NSDI 2011) Faircloud (Sigcomm 2012) Hadrian (NSDI 2013). Min-Guarantee. Each VM should be guaranteed a minimum bandwidth. (Oktopus et al.). Calculate VM Bandwidth. - PowerPoint PPT PresentationTRANSCRIPT
My Point of View about Bandwidth Sharing
Bin Wang
Outline
• Oktopus (Sigcomm 2011)
• TIVC (Sigcomm 2012)
• Seawall (NSDI 2011)
• Faircloud (Sigcomm 2012)
• Hadrian (NSDI 2013)
Min-Guarantee
• Each VM should be guaranteed a minimum bandwidth. (Oktopus et al.)
Calculate
VM Bandwidth VM Placement
Bulid Virtual Data Center
Oktopus [Hitesh Ballani et al. Sigcomm 2011]
• Virtual Cluster• <N, B>
• Virtual Oversubscribed Cluster
• <N, B, S, O>
BNBswitch
BNBgroup
O
BSBlink
O
BNBroot
TIVC [Di Xie et al. Sigcomm 2012]
• Temporally-interleaved Virtual Cluster
• Example: Single Peak <N, T, Bb, P>, where P=(T1, T2, B)
]T ,[T B
T] ,T [or ]T [0, BB(t)
21
21b
t
t
Network Proportionality
• The bandwidth allocated to a tenant should be proportional to its payment. (Seawall et al.)
• Per-flow allocation [B. Briscoe Sigcomm 2007]
• unfairness for jot flows
• Per-source allocation [Seawall Alan Shieh et al. NSDI 2011]
• asymmetric for bisection bandwidth allocation
• (similar to per-destination allocation)
High Utilization
• Spare network resources should be allocated to tenants with demand. (FairCloud et al.)
• Per-VM allocation [Gatekeeper H. Rodrigues et al. WIOV 2011]
• violate min-guarantee & proportionality• Per-SD allocation [FairCloud Lucian Popa et al. Sigcomm 2012]
[Hadrian Hitesh Ballani et al. NSDI 2013]
Good Allocation Strategies (1)
• Work conservation: As long as there is at least a tenant that has packets to send along link L, L cannot be idle. (FairCloud)
Good Allocation Strategies (2)
• Strategy-proofness: Tenants cannot improve their allocations by lying about their demands. (FairCloud)
Good Allocation Strategies (3)
• Utilization incentives: Tenants are never incentivized to reduce their actual demands on uncongested paths or to artificially leave links underutilized. (FairCloud)
ξC-AA 2 43
links. on two 22/3 bandwidth
get total will
ξ-C
A
links. on two 2/ bandwidth
get totalonly will
ξC
B
C-ξC 2 to2 from decrease n willutilizatio wholeThe
Good Allocation Strategies (4)
• Communication dependencies: A tenant’s communication dependency is a list of other tenants or peers that the tenant expects to communicate with. (Hadrian)
If, i) P: {Q}, ii) Q: {P, R}, iii) R: {*}, R cannot communicate with P.
Good Allocation Strategies (5)
• Min-guarantee: Total flows do get their minimum bandwidths. (Hadrian)
Good Allocation Strategies (6)
• Symmetry: The reverse allocation of each flow should match its original (forward) allocation. (FairCloud)
None of the state of art includes all the above issues.
• None of them is strategy-proofness because all of them are static allocations.
Hadrian
FairCloud
• PS-L:
• PS-P:
Strategy-proofness is requisite because it prevents malicious allocation actions.
My points of view
• Link incentives: Useful link will be work conservation as soon as possiable.
• Preferential policy: The last allocation statement, if triggered by newly allocations, should not be largely changed in a period.
• First-fit: The initial source&destination VMs through the link will acquire preferential policy.
• Other factors: Our proposal should not violate min-guarantee et al.
First Fit--Per-SD allocation
• Assume each VM has the same min-guarantee as 1.
00 , 1 LYXW YX
is a set of all VMs belonging to the proximate link l on first-fit period0L
Link l
p
q
r1
r2
r3
r4
For example, 10000
4321 rqrqrqrp WWWW
First Fit--Per-SD allocation
• When adding the transfer p'-r5, because it is also the first-fit, it's allocation weight:
Link l
p
q
r1
r2
r3
r4
P’ r5
10' 5
rpW
}or |,{ }1
1 ,1min{ 11
iii
iiYX LYLXYXW
i
The allocation strategy for newly D/S from the latest S/D.
is the number of the newly D/S from X at statement i.
i is the number of the decreased old S/D from X at statement i.
Instance
Link l
p
q
r1
r2
r3
r4
Link l
p
q
r1
r2
r3
r4
P’ r5
r6
10000
4321 rqrqrqrp WWWW
1 5.0}1 ,11
01min{
1
0'
1
0000
56
4321
rprq
rqrqrqrp
WW
WWWW
Instance (2)
Link l
p
q
r1
r2
r3
r4
P’ r5
r6
Link l
p
q
r1
r2
r3
r4
P’ r5
r6
r7
r8
p1
5.0}1 ,11
01min{
1}1 ,21
31min{
1
22
11
87
pr
rqrq
W
WW
Instance (3)
Link l
p
q
r1
r2
r3
r4
P’ r5
r6
}, W{WX-YLYLX Y-XX-Yii min :is of weight theso , and If
Link l
p
q
r1
r2
r3
r4
P’ r5
r6
r7
r8
0.5 is weight actural theSo
,1}1,31
31min{ e,perspectiv s' theFrom
,5.0}1 ,11
01min{ e,perspectiv s' theFrom
1
1
qp
qp
Wq
Wp
Disscussion
• The proposal is strategy-proofness.
Deeply increasing allocation does not affect the last allocation most.
• Deeply decreasing allocation will affect the benefit of the actor.
The strategy encourages the balance of the increasing&decreasing.
First-fit at Tenant Level
Link l
p
q
r1
r2
r3
r4
)} Z, ( ,or |,{ }1
1 ,1min{ i
i
iiYX LTZZTYTXYXW
10000
4321 rqrqrqrp WWWW
Link l
p
q
r1
r2
r3
r4
P’ r5
r6
5.0}1 ,11
01min{ 5.0}1 ,
11
01min{
1
1'
1
0000
56
4321
rprq
rqrqrqrp
WW
WWWW
First-fit at Tenant Level Case 1
}, W{WX-Y
LTTTYTX
Y-XX-Y
i
min :is of weight theso
, in element have and both , and If 2121
Link l
p
q
r1
r2
r3
r4
P’ r5
r6
Link l
p
q
r1
r2
r3
r4
P’ r5
r6
r7
r8
p1
0.5 is weight actural theSo
,1}1,31
31min{ e,perspectiv s' theFrom
,5.0}1 ,11
01min{ e,perspectiv s' theFrom
11
11
11
11
rp
rp
Wr
Wp
First-fit+Payment-guarantee
• The proportionality should represent VMs payment-guarantee. That means VMs with smaller minimal bandwidth should not acquire the profit fro
m VMs with larger one. (Hadrian)
Link l
p
q
r1
r2
r3
r4i
ii
1
1}
1
1 ,1min{ } ,min{
ii
rp
i
ir
ipi
W
BB
Proposal ComparisonWork conser
vationStrategy-proofness
Comm. depe.
Utilization incentives
Min-guarantee
Symmetry
Oktopus × × × × × ×Per-source √ × × × × ×Per-flow × × × × × ×
PS-L √ × × √ × √PS-P √ × × √ √ √
Hadrian √ × √ √ √ √
First-fit √ √ × √ x √
First-fit+payment
√ √ √ √ √ √
Future Work
• Consider the deployment in the tree-based topology/BCube
• Simulate on Estinet (compared with FairCloud, per-source, per-flow, Hadrian)
• Testbed (3 hops communication & Fat-tree)
Thanks