Load balance

Load balance
Course: Price of Anarchy
Professor: Michal Feldman
Student: Nave Frost
23/04/2014
Tight Bounds for Worst-Case Equilibria, Artur Czumaj and Berthold V¨ocking
Model
β€’ π‘š independent parallel links with speeds
𝑠1 , 𝑠2 , … , π‘ π‘š .
β€’ 𝑛 tasks with weights 𝑀1 , 𝑀2 , … , 𝑀𝑛 .
β€’ Goal: allocate tasks to links to minimize the
maximum load of the links in the system.
Definitions
β€’ Given a combination (𝑗1 , 𝑗2 , … , 𝑗𝑛 ) ∈ [π‘š]𝑛 of
pure strategies, one for each task:
β€’ The load of link 𝑗 is: π‘—π‘˜ =𝑗 𝑀𝑠 π‘˜.
𝑗
π‘€π‘˜
π‘—π‘˜ =𝑗𝑖 𝑠𝑗
𝑖
β€’ The cost for task 𝑖 is:
β€’ 𝑂𝑝𝑑 ≔
min
(𝑗1 ,𝑗2 ,…,𝑗𝑛
)∈[π‘š]𝑛
β€’ Observation: 𝑂𝑝𝑑 β‰₯
( max (
π‘—βˆˆ[π‘š]
max 𝑀𝑖
𝑖
max 𝑠𝑗
𝑗
.
𝑖:𝑗𝑖 =𝑗
𝑀𝑖
𝑠𝑗
))
Mixed strategies
β€’ Let 𝑝𝑖,𝑗 denote the probability that an task 𝑖
∈ 𝑛 sends the entire weight 𝑀𝑖 on link 𝑗.
β€’ Let 𝑙𝑗 denote the expected load on link j.
1
𝑙𝑗 ≔ 𝑠 π‘–βˆˆ 𝑛 𝑀𝑖𝑝𝑖,𝑗.
𝑗
β€’ Let 𝑐𝑖,𝑗 denote the expected cost for task 𝑖 on
link j.
𝑀𝑑 𝑝𝑑,𝑗
𝑀𝑖
𝑀𝑖
𝑐𝑖,𝑗 ≔
+
= 𝑙𝑗 + 1 βˆ’ 𝑝𝑖,𝑗
𝑠𝑗
𝑠𝑗
𝑠𝑗
𝑑≠𝑖
Nash equilibrium
β€’ Probabilities 𝑝𝑖,𝑗 define Nash equilibrium if
𝑝𝑖,𝑗 > 0 implies 𝑐𝑖,𝑗 ≀ 𝑐𝑖,π‘ž for every π‘ž ∈ [π‘š].
Price of anarchy
β€’ 𝑝𝑖,𝑗 - probabilities that define Nash equilibrium.
Define 𝐢𝑗 to be random variable indicating the
load of link 𝑗.
β€’ 𝐢𝑗 = 𝑠1
𝑗
β€’ 𝐽𝑖,𝑗
β€’
𝑛
𝑖=1 𝑀𝑖 𝐽𝑖,𝑗
1, 𝑀. 𝑝. 𝑝𝑖,𝑗
=
0,
𝑒𝑙𝑠𝑒
𝑛
E[𝐢𝑗 ] = 𝐸
𝑖=1 𝑀𝑖 𝐽𝑖,𝑗 =
1
1
𝑛
𝑛
𝑀
𝐸[𝐽
]
=
𝑀𝑝
𝑖,𝑗
𝑠𝑗 𝑖=1 𝑖
𝑠𝑗 𝑖=1 𝑖 𝑖,𝑗
1
𝑠𝑗
= 𝑙𝑗
Price of anarchy (2)
β€’ Let 𝑐 denote the maximum expected load over
all links, that is:
𝑐 = max 𝑙𝑗 = max 𝐸[𝐢𝑗 ]
π‘—βˆˆ[π‘š]
π‘—βˆˆ[π‘š]
β€’ Let the social cost 𝐢 denote the expected
maximum load over all links, that is:
𝐢 = 𝐸[ max 𝐢𝑗 ]
π‘—βˆˆ π‘š
β€’ Note that 𝑐 ≀ 𝐢 and possibly 𝑐 β‰ͺ 𝐢.
Price of anarchy (3)
β€’ The price of anarchy defined as :
𝐢
𝑅 = max 𝑂𝑝𝑑
𝑝𝑖,𝑗
Claim 1
β€’ Claim 1:
In Nash equilibrium, if 𝑝𝑖,π‘ž > 0 for certain task 𝑖 ∈ 𝑛 and link π‘ž ∈ π‘š , then
for every link j ∈ [m]
𝑙𝑗 + 𝑀𝑠 𝑖 β‰₯ π‘™π‘ž
𝑗
In particular, if 𝑙𝑗 + 1 < π‘™π‘ž then 𝑀𝑖 > 𝑠𝑗 .
β€’ Proof:
𝑐𝑖,𝑗 = 𝑙𝑗 + 1 βˆ’ 𝑝𝑖,𝑗
𝑐𝑖,π‘ž = π‘™π‘ž + 1 βˆ’
𝑀𝑖
𝑀𝑖
≀
𝑙
+
𝑗
𝑠𝑗
𝑠𝑗
𝑖 β‰₯ 𝑙
𝑝𝑖,π‘ž 𝑀
π‘ž
π‘ π‘ž
Since 𝑝𝑖,π‘ž > 0, the
definition of Nash equilibria implies that 𝑐𝑖,π‘ž ≀ 𝑐𝑖,𝑗 and
𝑀𝑖
hence, π‘™π‘ž ≀ 𝑙𝑗 + .
𝑠𝑗
β€’ The second part of the claim follows trivially from the first one.
Identical links
β€’ In the case of identical links all speeds are
identical, i.e., for all 𝑗1 , 𝑗2 ∈ π‘š
𝑠𝑗1 = 𝑠𝑗2
Identical links (2)
β€’ Theorem:
For π‘š identical links the price of anarchy is
𝑂
log π‘š
log log π‘š
.
β€’ Proof:
Let us first re-scale the weights and speeds in the
problem and assume, without loss of generality,
that for all j ∈ π‘š : 𝑠𝑗 = 1 and 𝑂𝑝𝑑 = 1.
Identical links (3)
β€’ Since 𝑂𝑝𝑑 = 1 and we observed that 𝑂𝑝𝑑
max 𝑀𝑖
β‰₯ 𝑖
, then βˆ€i ∈ 𝑛 : 𝑀𝑖 ≀ 1.
max 𝑠𝑗
𝑗
β€’ Since 𝑂𝑝𝑑 = 1 there must exist a 𝑗 such that
𝑙𝑗 ≀ 1.
β€’ Since 𝑝𝑖,𝑗 define Nash equilibrium it implies
that, according to claim 1:
𝑐 = max 𝑙𝑗 ≀ 2
π‘—βˆˆ[π‘š]
Identical links (4)
β€’ We will estimate the load 𝐢𝑗 of any link 𝑗 ∈ [π‘š] by applying to 𝐢𝑗 =
𝑛
𝑖=1 𝑀𝑖 𝐽𝑖,𝑗 a standard concentration inequality due to Hoeffding
bound.
β€’ Hoeffding bound: Let 𝑋1 , … , 𝑋𝑁 be independent random variables
with values in the interval [0, 𝑧] for some 𝑧 > 0, and let 𝑑𝑋 =
𝑛
𝑖=1 𝑋𝑖 ,
then for any 𝑑 it holds that Pr 𝑋 β‰₯ 𝑑 ≀
β€’ Pr 𝐢𝑗 β‰₯ 𝑑 = Pr
𝑒𝑙𝑗 𝑑
𝑑
≀
2𝑒 𝑑
𝑑
𝑛
𝑖=1 𝑀𝑖 𝐽𝑖,𝑗
β‰₯𝑑 ≀
For any 𝑑 > 0.
𝑒𝐸
𝑛
𝑖=1 𝑀𝑖 𝐽𝑖,𝑗
𝑑
𝑒𝐸 𝑋
𝑑
𝑑
=
𝑧
𝑒𝐸 𝐢𝑗
𝑑
𝑑
=
Identical links (5)
2𝑒 𝑑
We saw that Pr 𝐢𝑗 β‰₯ 𝑑 ≀
For any 𝑑 > 0.
𝑑
3 ln π‘š
Therefore, if we pick 𝑑 =
we will get the following:
ln ln π‘š
∞
C = 𝐸 max 𝐢𝑗 ≀ t+
π‘—βˆˆ π‘š
∞
π‘š·πœ 2𝑒 𝜏
≀ t+
𝜏=𝑑
Pr βˆƒπ‘— ∈ π‘š : 𝐢𝑗 β‰₯ 𝜏 ·πœ
𝜏=𝑑
𝜏
∞
2βˆ’πœ
≀ t+
𝜏=𝑑
3 ln π‘š
≀𝑑+1=
+1
ln ln π‘š
Upper bound – Speed varies
β€’ Lemma 1:
The maximum expected load 𝑐 satisfies
c = 𝑂𝑝𝑑·π‘‚ min
log π‘š
𝑠1
, log
log log π‘š
π‘ π‘š
where it is assumed that the speeds satisfy 𝑠1 β‰₯ β‹― β‰₯ π‘ π‘š .
β€’ Lemma 2:
The social cost 𝐢 satisfies
log π‘š
C = 𝑂𝑝𝑑·π‘‚
+1
𝑂𝑝𝑑· log π‘š
log
𝑐
Lemma 1 - Proof
β€’ 𝑝𝑖,𝑗 define a Nash equilibrium.
β€’ Without loss of generality, assume 𝑠1 β‰₯ β‹―
β‰₯ π‘ π‘š .
β€’ Scale weights of such that 𝑂𝑝𝑑 = 1.
β€’ Will show:
c=O
log π‘š
log log π‘š
and c = O log
𝑠1
π‘ π‘š
c=O
log π‘š
log log π‘š
– Proof
Proof:
For π‘˜ β‰₯ 1, define π‘—π‘˜ to be the smallest index in
{0,1, … , π‘š} such that π‘™π‘—π‘˜ +1 < π‘˜ or, if no such index
exists, π‘—π‘˜ = π‘š.
Observe:
β€’ for every π‘˜ β‰₯ 1 with 0 < π‘—π‘˜ , all links j < π‘—π‘˜ have
load at least π‘˜.
β€’ for every π‘˜ β‰₯ 1 with π‘—π‘˜ < π‘š, link π‘—π‘˜ + 1 has load
less than π‘˜.
c=O
log π‘š
log log π‘š
– Proof (2)
β€’ Let 𝑐 βˆ— = 𝑐 βˆ’ 1 .
β€’ Will show that 𝑗1 β‰₯ 𝑐 βˆ— !
β€’ Hence, (recall that 𝑗1 ≀ π‘š)
𝑐≀Γ
βˆ’1
π‘š +1=
log π‘š
log log π‘š
1+π‘œ 1
β€’ Gamma function: Ξ“ N + 1 = N! and Ξ“ βˆ’1 𝑁 is
the inverse.
Known:
log 𝑁
βˆ’1
Ξ“
𝑁 =
(1 + π‘œ(1))
log log 𝑁
c=O
β€’
log π‘š
log log π‘š
– Proof (3)
Claim: 𝑗𝑐 βˆ— > 0 and hence 𝑙1 β‰₯ 𝑐 βˆ— .
β€’ Proof:
For the purpose of contradiction, assume:
𝑙1 < 𝑐 βˆ— ≀ 𝑐 βˆ’ 1.
Let π‘ž denote the link with the maximum expected load.
π‘™π‘ž = 𝑐 > 𝑙1 + 1
If there exist a task 𝑖 with weight 𝑀𝑖 ≀ 𝑠1 :
𝑖 ≀ 𝑙 + 𝑀𝑖 ≀ 𝑙 + 1 < 𝑙
𝑐𝑖,1 = 𝑙1 + 1 βˆ’ 𝑝𝑖,1 𝑀
1
1
π‘ž
𝑠1
𝑠1
which contradicts to Claim 1.
Hence, there is no task 𝑖 with weight 𝑀𝑖 ≀ 𝑠1 and since:
𝑂𝑝𝑑 β‰₯
max 𝑀𝑖
𝑖
max 𝑠𝑗
𝑗
It leads to contradiction.
=
max 𝑀𝑖
𝑖
𝑠1
>1
c=O
log π‘š
log log π‘š
– Proof (4)
β€’ Claim: For π‘˜ β‰₯ 1, π‘—π‘˜ β‰₯ (π‘˜ + 1)π‘—π‘˜+1 .
β€’ Proof:
Let 𝑇 be the set of tasks in the system that have
positive probability on at least one of the links
in {1, . . . , π‘—π‘˜+1 }.
Fix an optimal allocation strategy π‘‚π‘ƒπ‘‡π‘†π‘‘π‘Ÿ. We
distinguish between two different ways of how
π‘‚π‘ƒπ‘‡π‘†π‘‘π‘Ÿ might allocate the tasks in 𝑇 to the links.
c=O
log π‘š
log log π‘š
– Proof (5)
β€’ Case 1:
For the purpose of contradiction, assume π‘‚π‘ƒπ‘‡π‘†π‘‘π‘Ÿ allocates at
least one of the tasks in 𝑇 to a link π‘ž, π‘ž > π‘—π‘˜ .
π‘Šπ‘‡ denote the minimum weight of the tasks in 𝑇.
For all j ∈ 1, . . . , π‘—π‘˜+1 :
𝑙𝑗 β‰₯ π‘˜ + 1
π‘™π‘—π‘˜ +1 < π‘˜
Therefore, according to claim 1 π‘Šπ‘‡ > π‘ π‘—π‘˜ +1 .
But this implies that allocating a task from 𝑇 to link π‘ž gives
π‘Šπ‘‡
π‘Šπ‘‡
cost at least
β‰₯
> 1.
π‘ π‘ž
π‘ π‘—π‘˜ +1
This means 𝑂𝑝𝑑 > 1 and hence the contradiction.
c=O
log π‘š
log log π‘š
– Proof (6)
β€’ Case 2:
Now let us assume π‘‚π‘ƒπ‘‡π‘†π‘‘π‘Ÿ allocates all tasks in 𝑇 to the links {1, … , π‘—π‘˜ }.
We will show that this implies π‘—π‘˜ β‰₯ (π‘˜ + 1)π‘—π‘˜+1 .
π‘ŠΞ£T denote the sum of the weights of the tasks in 𝑇.
π‘—π‘˜+1
π‘ŠΞ£T =
𝑀𝑖 β‰₯
π‘–βˆˆπ‘‡
𝑀𝑖
π‘–βˆˆπ‘‡
π‘—π‘˜+1
π‘—π‘˜+1 𝑛
𝑝𝑖,𝑗 =
𝑗=1
𝑀𝑖 𝑝𝑖,𝑗 =
𝑗=1 π‘–βˆˆπ‘‡
π‘—π‘˜+1
𝑀𝑖 𝑝𝑖,𝑗 =
𝑗=1 𝑖=1
π‘—π‘˜+1
𝑙𝑗 𝑠𝑗 β‰₯ (π‘˜ + 1)
𝑗=1
On the other hand:
π‘—π‘˜
π‘ŠΞ£T ≀
𝑠𝑗
𝑗=1
Hence:
π‘—π‘˜
π‘—π‘˜+1
𝑠𝑗 β‰₯ (π‘˜ + 1)
𝑗=1
𝑠𝑗
𝑗=1
Since the sequence of link speeds is non-increasing, this implies that π‘—π‘˜ β‰₯ (π‘˜ + 1)π‘—π‘˜+1 .
𝑠𝑗
𝑗=1
c=O
log π‘š
log log π‘š
– Proof (7)
We combine both proven claims:
β€’ Claim: 𝑗𝑐 βˆ— > 0 and hence 𝑙1 β‰₯ 𝑐 βˆ— .
β€’ Claim: For π‘˜ β‰₯ 1, π‘—π‘˜ β‰₯ (π‘˜ + 1)π‘—π‘˜+1 .
And get:
𝑗1 β‰₯ 𝑐 βˆ— ! 𝑗𝑐 βˆ— β‰₯ 𝑐 βˆ— !
As wanted.
Hence, 𝑐 = O
log π‘š
log log π‘š
.
c = O log
β€’ c = O log
𝑠1
π‘ π‘š
– Proof
𝑠1
π‘ π‘š
β€’ Proof:
We will claim that the speeds of the links
𝑗1 , 𝑗2 , … increase in a geometric fashion.
c = O log
𝑠1
π‘ π‘š
– Proof (2)
β€’ Claim For 1 ≀ π‘˜ ≀ 𝑐 βˆ’ 3, π‘ π‘—π‘˜+2 +1 β‰₯ 2π‘ π‘—π‘˜ +1 .
β€’ Proof:
Fix an optimal strategy π‘‚π‘ƒπ‘‡π‘†π‘‘π‘Ÿ.
For every link 𝑗 β€² ≀ π‘—π‘˜+2 :
𝑙𝑗 β€² β‰₯ π‘˜ + 2 > 1 = 𝑂𝑝𝑑
Hence, There is task 𝑖 and link 𝑗 ∈ {1, … , π‘—π‘˜+2 } with 𝑝𝑖,𝑗 > 0
that π‘‚π‘ƒπ‘‡π‘†π‘‘π‘Ÿ allocate to link in {π‘—π‘˜+2 + 1, … , π‘š}.
Note that:
𝑀𝑖 ≀ π‘ π‘—π‘˜+2 +1
Therefore, there exists a link j ∈ {1, … , π‘—π‘˜+2 } and a task 𝑖 of
weight 𝑀𝑖 ≀ π‘ π‘—π‘˜+2+1 with 𝑝𝑖,𝑗 > 0.
c = O log
𝑠1
π‘ π‘š
On the one hand:
𝑐𝑖,𝑗 = 𝑙𝑗 + 1 βˆ’ 𝑝𝑖,𝑗
On the other hand:
𝑐𝑖,π‘—π‘˜ +1 = π‘™π‘—π‘˜ +1 + 1 βˆ’ 𝑝𝑖,π‘—π‘˜ +1
𝑀𝑖
– Proof (3)
𝑀𝑖
β‰₯ 𝑙𝑗 β‰₯ π‘˜ + 2
𝑠𝑗
≀ π‘™π‘—π‘˜ +1 +
𝑀𝑖
<π‘˜+
𝑀𝑖
π‘ π‘—π‘˜ +1
π‘ π‘—π‘˜ +1
π‘ π‘—π‘˜+1
Since we assume the system is in a Nash equilibrium, and 𝑝𝑖,𝑗 > 0 the
cost of task 𝑖 on link 𝑗 cannot be larger than the cost of task 𝑖 on link π‘—π‘˜
π‘ π‘—π‘˜+2 +1
𝑀𝑖
π‘˜+2β‰€π‘˜+
β‰€π‘˜+
π‘ π‘—π‘˜ +1
π‘ π‘—π‘˜+1
Clearly, this inequality implies that:
2π‘ π‘—π‘˜+1 ≀ π‘ π‘—π‘˜+2 +1
Hence, the claim is shown.
c = O log
𝑠1
π‘ π‘š
– Proof (4)
The previous claim says that in a Nash
equilibrium the speeds increase geometrically
with the expected load.
This implies that:
π‘ π‘š ≀ 𝑠𝑗1+1 ≀
Thus,
π‘βˆ’2
βˆ’
2 2
· π‘ π‘—π‘βˆ’1+1 ≀
π‘βˆ’2
βˆ’
2 2
𝑠1
c ≀ 2 log
+ O(1)
π‘ π‘š
· 𝑠1
Lemma 2 - Proof
We wish to show that 𝐢 satisfies:
log π‘š
C = 𝑂𝑝𝑑·π‘‚
+1
𝑂𝑝𝑑· log π‘š
log
𝑐
Our goal is to show, for every link j ∈ [π‘š], it is
unlikely that 𝐢𝑗 deviates much from its
expectation.
For this purpose, we will use a Hoeffding bound.
Lemma 2 – Proof (2)
β€’ Claim For every task 𝑖 and link 𝑗 with 𝑝𝑖,𝑗
1
∈ 0, 4 , 𝑀𝑖 ≀ 12·π‘ π‘— ·π‘‚𝑝𝑑
β€’ Proof:
We extend the definition of π‘—π‘˜ to hold for
arbitrary 𝑂𝑝𝑑 in natural way: for π‘˜ β‰₯ 1, we
define π‘—π‘˜ as the smallest index in {1, … , π‘š} such
that π‘™π‘—π‘˜ +1 < π‘˜·π‘‚𝑝𝑑, or, π‘—π‘˜ = π‘š if no such index
exists.
Lemma 2 – Proof (3)
Let’s look at link 𝑗 ∈ {π‘—π‘˜ + 1, … , π‘—π‘˜βˆ’1 } for some π‘˜ ≀ 𝑐 βˆ’ 3.
On the one hand:
𝑖 β‰₯ π‘˜ βˆ’ 1 ·π‘‚𝑝𝑑 + 3𝑀𝑖
𝑐𝑖,𝑗 = 𝑙𝑗 + 1 βˆ’ 𝑝𝑖,𝑗 𝑀
𝑠
4𝑠
𝑗
On the other hand:
𝑐𝑖,π‘—π‘˜+2 +1 = π‘™π‘—π‘˜+2 +1 + 1 βˆ’ 𝑝𝑖,π‘—π‘˜+2 +1
𝑀𝑖
≀ π‘˜ + 2 ·π‘‚𝑝𝑑 + 2𝑠
𝑗
𝑀𝑖
π‘ π‘—π‘˜+2 +1
≀ π‘™π‘—π‘˜+2 +1 + 𝑠
𝑀𝑖
π‘—π‘˜+2 +1
𝑗
Since we assume the system is in a Nash equilibrium, and 𝑝𝑖,𝑗 > 0 the
cost of task 𝑖 on link 𝑗 cannot be larger than the cost of task 𝑖 on link
π‘—π‘˜+2 + 1.
𝑖 ≀ π‘˜ + 2 ·π‘‚𝑝𝑑 + 𝑀𝑖 . Which imply :
Hence, π‘˜ βˆ’ 1 ·π‘‚𝑝𝑑 + 3𝑀
4𝑠
2𝑠
𝑗
𝑗
𝑀𝑖 ≀ 12·π‘‚𝑝𝑑 ·π‘ π‘—
Lemma 2 – Proof (4)
It remains to investigate the case 𝑗 ≀ π‘—π‘˜ , where π‘˜ = 𝑐 βˆ’ 3 .
On the one hand:
𝑀𝑖
𝑐𝑖,1 = 𝑙1 + 1 βˆ’ 𝑝𝑖,1
≀ 𝑐·π‘‚𝑝𝑑 + 𝑂𝑝𝑑 = 𝑐 + 1 ·π‘‚𝑝𝑑
𝑠1
On the other hand:
𝑀𝑖
3𝑀𝑖
3𝑀𝑖
𝑐𝑖,𝑗 = 𝑙𝑗 + 1 βˆ’ 𝑝𝑖,𝑗
β‰₯ π‘˜·π‘‚𝑝𝑑 +
β‰₯ 𝑐 βˆ’ 4 ·π‘‚𝑝𝑑 +
𝑠𝑗
4𝑠𝑗
4𝑠𝑗
Since we assume the system is in a Nash equilibrium, and 𝑝𝑖,𝑗 > 0 the cost of
task 𝑖 on link 𝑗 cannot be larger than the cost of task 𝑖 on link 1.
Hence, 𝑐 βˆ’ 4 ·π‘‚𝑝𝑑 +
This proves the claim.
3𝑀𝑖
4𝑠𝑗
≀ 𝑐 + 1 ·π‘‚𝑝𝑑. Which imply :
20
𝑀𝑖 ≀
·π‘‚𝑝𝑑·π‘ π‘— < 12·π‘‚𝑝𝑑·π‘ π‘—
3
Lemma 2 – Proof (5)
Now let us focus on a single link 𝑗 ∈ [π‘š].
We apply the claim in order to show that it is unlikely that 𝐢𝑗 deviates
much from its expectation.
Denote:
𝑇𝑗,1 - the set of tasks with 𝑝𝑖,𝑗 ∈ 0, 14 .
𝑇𝑗,2 - the set of tasks with 𝑝𝑖,𝑗 ∈ 14, 1 .
𝐢𝑗,1 - random variable that describe the cost on link 𝑗 only counting
tasks in 𝑇𝑗,1 .
𝐢𝑗,1 - random variable that describe the cost on link 𝑗 only counting
tasks in 𝑇𝑗,2 .
Since only tasks with 𝑝𝑖,𝑗 > 0 can be allocated to link 𝑗:
𝐢𝑗 = 𝐢𝑗,1 + 𝐢𝑗,2
Lemma 2 – Proof (6)
Let us consider only the tasks in 𝑇𝑗,1 .
𝐢𝑗,1 = 𝑠1
𝑀𝑖 𝐽𝑖,𝑗
𝑗
π‘–βˆˆπ‘‡π‘—,1
β€’ Remainder :
Hoeffding bound: Let 𝑋1 , … , 𝑋𝑁 be independent random variables with values
in the interval [0, 𝑧] for some 𝑧 > 0, and let 𝑋 = 𝑛𝑖=1 𝑋𝑖 , then for any 𝑑 it
𝑒𝐸 𝑋 𝑑 𝑧
holds that Pr 𝑋 β‰₯ 𝑑 ≀ (
)
𝑑
According to the previous claim:
𝑖 ≀ 12·π‘‚𝑝𝑑
max 𝑀
𝑠
π‘–βˆˆπ‘‡π‘—,1 𝑗
·
𝛼 𝑂𝑝𝑑
Pr 𝐢𝑗,1 β‰₯ 𝛼·π‘‚𝑝𝑑 ≀
For any 𝛼 > 1.
𝑒𝐸 𝐢𝑗,1
𝛼·π‘‚𝑝𝑑
·
12 𝑂𝑝𝑑
≀
𝑒𝑐
𝛼·π‘‚𝑝𝑑
𝛼
12
Lemma 2 – Proof (7)
Let us consider only the tasks in 𝑇𝑗,2 .
𝐢𝑗,2 = 𝑠1
𝑀𝑖 𝐽𝑖,𝑗 ≀ 𝑠1
𝑗
π‘–βˆˆπ‘‡π‘—,2
𝐸 𝐢𝑗,2 = 𝐸
1
𝑠𝑗
π‘–βˆˆπ‘‡π‘—,2
𝑀𝑖 𝐽𝑖,𝑗 = 𝑠1
𝑀𝑖 𝐸 𝐽𝑖,𝑗 > 4𝑠1
𝑗
π‘–βˆˆπ‘‡π‘—,2
𝑀𝑖
𝑗
π‘–βˆˆπ‘‡π‘—,2
π‘–βˆˆπ‘‡π‘—,2
Hence, we get:
𝐢𝑗,2 < 4𝐸 𝐢𝑗,2 ≀ 4𝑐
Hence, for every 𝛼 > 1
𝑒𝑐
Pr 𝐢𝑗 β‰₯ 4𝑐 + 𝛼·π‘‚𝑝𝑑 ≀
𝛼·π‘‚𝑝𝑑
𝑀𝑖
𝑗
𝛼
12
Lemma 2 – Proof (8)
Denote β„’ = max 𝐢𝑗.
π‘—βˆˆ[π‘š]
𝑒𝑐
Pr β„’ β‰₯ 4𝑐 + 𝛼·π‘‚𝑝𝑑 ≀ π‘š
𝛼·π‘‚𝑝𝑑
Since of union bound.
𝛼
12
Lemma 2 – Proof (9)
Recall that 𝐢 is defined to be 𝐸 β„’ .
Hence, for every πœ† > 1, we can estimate 𝐢 as follows.
∞
𝐢 = 𝐸 β„’ ≀ 4𝑐 + Ξ»·π‘‚𝑝𝑑 + 𝑂𝑝𝑑·
Pr β„’ β‰₯ 4𝑐 + πœ† + 𝑑 ·π‘‚𝑝𝑑 𝑑𝑑
0
∞
≀ 4𝑐 + Ξ»·π‘‚𝑝𝑑 + 𝑂𝑝𝑑·π‘š·
0
𝑒𝑐
≀ 4𝑐 + Ξ»·π‘‚𝑝𝑑 + 𝑂𝑝𝑑·π‘š·
πœ†·π‘‚𝑝𝑑
Notice that for Ξ» β‰₯
21𝑒𝑐
𝑂𝑝𝑑
we have
∞
𝑒𝑐
0
πœ†·π‘‚𝑝𝑑
𝑑
12
πœ†+𝑑
𝑒𝑐
πœ† + 𝑑 ·π‘‚𝑝𝑑
πœ†
∞
12
𝑑𝑑 =
·
0
𝑑𝑑
·
𝑑
𝑒𝑐
πœ†·π‘‚𝑝𝑑
12
ln
12
πœ† 𝑂𝑝𝑑
𝑒𝑐
12
𝑑𝑑
≀ 4 and therefore we obtain:
𝑒𝑐
𝐢 ≀ 4𝑐 + Ξ»·π‘‚𝑝𝑑 + 4·π‘‚𝑝𝑑·π‘š·
πœ†·π‘‚𝑝𝑑
πœ†
12
Lemma 2 – Proof (10)
For Ξ» β‰₯
𝑐
(2Ξ“ βˆ’1
𝑂𝑝𝑑
(𝑒·π‘š
·
12 𝑂𝑝𝑑
𝑐
)) then:
𝑒𝑐
πœ†·π‘‚𝑝𝑑
To prove this inequality, let us set:
πœ†0 =
𝑐
1+Ξ“
𝑂𝑝𝑑
πœ†
12
≀
βˆ’1
1
π‘š
(𝑒·π‘š
12·π‘‚𝑝𝑑
𝑐)
Observe that for any integer π‘˜ β‰₯ 1
(π‘˜ 𝑒)π‘˜ β‰₯ π‘˜ βˆ’ 1 !
Therefore
πœ†0 ·π‘‚𝑝𝑑
𝑒·π‘
πœ†0 ·π‘‚𝑝𝑑
𝑐
β‰₯
πœ†0 ·π‘‚𝑝𝑑
βˆ’ 1 ! ·π‘’ βˆ’1
𝑐
Since πœ†0 ensures that:
πœ†0 ·π‘‚𝑝𝑑
12·π‘‚𝑝𝑑
𝑐
βˆ’ 1 ! ·π‘’ βˆ’1 β‰₯ π‘š
𝑐
We can conclude that for any πœ† β‰₯ πœ†0 we have:
𝑐
12·π‘‚𝑝𝑑
πœ†0
πœ†0 ·π‘‚𝑝𝑑
πœ†
𝑐
12
12
πœ†·π‘‚𝑝𝑑
πœ†0 ·π‘‚𝑝𝑑
πœ†0 ·π‘‚𝑝𝑑
12·π‘‚𝑝𝑑
𝑐
β‰₯
=
β‰₯ π‘š
𝑒·π‘
𝑒·π‘
𝑒·π‘
𝑐
12·π‘‚𝑝𝑑
=π‘š
Lemma 2 – Proof (11)
Therefore, if we set:
πœ† = max
21𝑒𝑐
𝑐
1,
,
𝑂𝑝𝑑 𝑂𝑝𝑑
2Ξ“
βˆ’1
𝑒·π‘š
·
12 𝑂𝑝𝑑
𝑐
=𝑂 1
𝑐
+
𝑂𝑝𝑑
Upper bound - Proof
In Lemma 2 we saw
𝑐
log π‘š
𝐢 = 𝑂𝑝𝑑·π‘‚ 1 +
+
𝑂𝑝𝑑 log 𝑂𝑝𝑑· log
π‘š
𝑐
If 𝑐 ≀ 𝑂𝑝𝑑 then
C = 𝑂𝑝𝑑·π‘‚
log π‘š
+1
𝑂𝑝𝑑· log π‘š
log
𝑐
If 𝑐 > 𝑂𝑝𝑑 then by Lemma 1
𝑐
log π‘š
log π‘š
=𝑂
=𝑂
𝑂𝑝𝑑
log log π‘š
log 𝑂𝑝𝑑· 𝑐log π‘š
Hence,
𝑐
log π‘š
log π‘š
C = 𝑂𝑝𝑑·π‘‚ 1 +
+
= 𝑂𝑝𝑑·π‘‚
+1
𝑂𝑝𝑑 log 𝑂𝑝𝑑· log π‘š
log 𝑂𝑝𝑑· log π‘š
𝑐
𝑐
Lower bound
Let’s assume without loss of generality that π‘š is an integer.
We consider 𝐾 + 1 groups of links 𝐸0 , 𝐸1 , … , 𝐸𝐾 .
And 𝐾 + 1 groups of tasks 𝑇0 , 𝑇1 , … , 𝑇𝐾 .
For all 0 ≀ π‘˜ ≀ 𝐾
𝐾!
πΈπ‘˜ = · π‘š
π‘˜!
π‘‡π‘˜ = π‘˜· πΈπ‘˜
For all 𝑗 ∈ πΈπ‘˜
For all 𝑖 ∈ π‘‡π‘˜
𝑠𝑗 = 2π‘˜
𝑀𝑖 = 2π‘˜
Will consider a pure strategy 𝑆 in which for each 𝑗 ∈ πΈπ‘˜ it assigns exactly π‘˜ tasks from
π‘‡π‘˜ with probability 1 to be allocated to 𝑗.
Lower bound – (2)
In our construction 𝐾 can be chosen to be any positive integer that satisfies:
𝐾
πΈπ‘˜ ≀ π‘š
π‘˜=0
Note that:
𝐾
𝐾
πΈπ‘˜ =
π‘˜=0
π‘˜=0
π‘š·πΎ!
= π‘š·πΎ! ·
π‘˜!
𝐾
π‘˜=0
1
≀ π‘š·πΎ! ·π‘’
π‘˜!
Thus, our analysis can be carried over for all 𝐾 satisfying:
π‘š·πΎ! ·π‘’ ≀ π‘š
And hence for all 𝐾 satisfying:
π‘š
log π‘š
𝐾 ≀ Ξ“ βˆ’1
βˆ’1=𝛩
𝑒
log log π‘š
Lower bound – (3)
β€’ Claim Strategy 𝑆 satisfies the following
properties:
1. The maximum load is c = 𝐾.
2. The social optimum is 1 ≀ 𝑂𝑝𝑑 ≀ 2.
3. The system is in Nash equilibrium.
β€’ Proof:
1. For all 𝑗 ∈ πΈπ‘˜ the load 𝐢𝑗 =
π‘˜ ·2 π‘˜
𝑠𝑗
=
π‘˜·2π‘˜
2π‘˜
= π‘˜.
Lower bound – (4)
2. First we will observe that:
max 𝑀𝑖
2π‘˜
𝑂𝑝𝑑 β‰₯
= π‘˜=1
max 𝑠𝑗
2
𝑖
𝑗
Let’s consider a strategy in which for 0 ≀ π‘˜ < 𝐾 for each 𝑗 ∈ πΈπ‘˜ it assigns a single task from π‘‡π‘˜+1 with
probability 1 to be allocated to 𝑗.
Notice that:
𝐾!
πΈπ‘˜ = · π‘š
π‘˜!
𝐾!
𝐾!
π‘‡π‘˜+1 = (π‘˜+1)· πΈπ‘˜+1 = (π‘˜+1)·
· π‘š= · π‘š
π‘˜+1 !
π‘˜!
For all 𝑗 ∈ πΈπ‘˜
𝑠𝑗 = 2π‘˜
For all 𝑖 ∈ π‘‡π‘˜+1
𝑀𝑖 = 2π‘˜+1
The cost of every link in the system is at most 2.
𝑂𝑝𝑑 ≀ 2
Lower bound – (5)
3. Let us take any task 𝑖 that is allocated to link 𝑗 ∈ πΈπ‘˜ , for π‘˜
β‰₯ 1.
Lets see that it is not beneficial for 𝑖 to move to link π‘Ÿ ∈ 𝐸𝑑 .
𝑀𝑖
𝑐𝑖,𝑗 = 𝑙𝑗 + 1 βˆ’ 𝑝𝑖,𝑗
= 𝑙𝑗 = π‘˜
𝑠𝑗
𝑀𝑖
2π‘˜
𝑐𝑖,π‘Ÿ = π‘™π‘Ÿ + 1 βˆ’ 𝑝𝑖,π‘Ÿ
= π‘™π‘Ÿ + 𝑑 = 𝑑 + 2π‘˜βˆ’π‘‘
π‘ π‘Ÿ
2
Since 𝑑 + 2π‘˜βˆ’π‘‘ β‰₯ π‘˜ for any non-negative 𝑑 and π‘˜, none of the
tasks allocated to 𝑗 has an incentive to migrate to another link.
Therefore, the system is in a Nash equilibrium.
Lower bound – Mixed strategy
We will slightly modify the pure strategy 𝑆 to
obtain a mixed strategy 𝑆′ for which:
log π‘š
𝐢 = 𝑐·π›©
log(log(π‘š) 𝑐)
For 0 ≀ π‘˜ < 𝐾, 𝑆′ will allocate tasks in the same
manner 𝑆 did.
1
For all 𝑖 ∈ 𝑇𝐾 and 𝑗 ∈ 𝐸𝐾 , 𝑆′ will set 𝑝𝑖,𝑗 = π‘š.
Lower bound – Mixed strategy – (2)
β€’ Claim Strategy 𝑆′ satisfies the following properties:
1. The maximum load is c = 𝐾.
2. The social optimum is 1 ≀ 𝑂𝑝𝑑 ≀ 2.
3. The system is in Nash equilibrium.
4. The social cost is 𝐢 = Ω
log π‘š
log(log(π‘š) 𝐾)
β€’ Proof:
1. The maximum load 𝑐 is the same as for strategy 𝑆′.
Lower bound – Mixed strategy – (3)
𝟐. The value of opt is unaffected by the modification of the probabilities.
πŸ‘. We have to show that for 𝑖 ∈ 𝑇𝐾 it is not beneficial to move to link π‘Ÿ ∈ 𝐸𝑑 .
For 𝑗 ∈ 𝐸𝐾
𝑀𝑖
1 2𝐾
1
𝑐𝑖,𝑗 = 𝑙𝑗 + 1 βˆ’ 𝑝𝑖,𝑗
=𝐾+ 1βˆ’
=
𝐾
+
1
βˆ’
𝑠𝑗
π‘š 2𝐾
π‘š
We can see that the costs for task 𝑖 on link 𝑗 slightly increased.
𝑀𝑖
2𝐾
𝑐𝑖,π‘Ÿ = π‘™π‘Ÿ + 1 βˆ’ 𝑝𝑖,π‘Ÿ
= π‘™π‘Ÿ + 𝑑 = 𝑑 + 2πΎβˆ’π‘‘
π‘ π‘Ÿ
2
The cost for task 𝑖 on link π‘Ÿ ∈ 𝐸𝑑 remains the same as before.
1
Since 𝑑 + 2πΎβˆ’π‘‘ β‰₯ 𝐾 + 1 β‰₯ 𝐾 + 1 βˆ’
for any non-negative 𝑑 and 𝐾, none
π‘š
of the tasks allocated to 𝑗 has an incentive to migrate to another link.
Therefore, the system is in a Nash equilibrium.
Lower bound – Mixed strategy – (4)
πŸ’. To observe this property, we notice that the allocation of the tasks
in 𝑖 ∈ 𝑇𝐾 to the links in 𝐸𝐾 corresponds to the allocation problem of
throwing π‘š·πΎ balls uniformly at random into π‘š bins.
In expectation, it is known that the expected maximum occupancy in
this allocation problem is:
log π‘š
𝛩 𝐾+
log(log(π‘š) 𝐾)
Since 𝐾 = 𝛩
log π‘š
log log π‘š
in our case. We get:
log π‘š
𝛩
log(log(π‘š) 𝐾)
Since for 𝑖 ∈ 𝑇𝐾 and j ∈ 𝐸𝐾 , 𝑀𝑖 = 𝑠𝑗 , this bound on the maximum
occupancy directly implies a lower bound on the social cost.