Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model Magn´ us M. Halld´ orsson and Pradipta Mitra
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
1 / 20
Wireless communication
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
2 / 20
Path loss
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
3 / 20
Modeling interference
“Links” are intended transmissions. How do we capture interference between links?
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
4 / 20
Modeling interference
Physical model of signal transmission Signal received at r from s =
Halld´ orsson and Mitra ()
P d(s,r )α ,
if s transmits with power P.
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
4 / 20
Modeling interference (contd.)
Physical model of interference Pv d(sv ,rv )α Pu d(su ,rv )α
(Good) Signal at rv = Interference at rv = Ambient noise N
Signal to Interference + Noise Ratio SINR =
Pv d(sv ,rv )α Pu +N d(su ,rv )α
Link succeeds iff SINR ≥ β. Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
5 / 20
Modeling interference (contd.)
Physical model of interference Pv d(sv ,rv )α Pu d(su ,rv )α
(Good) Signal at rv = Interference at rv = Ambient noise N
Signal to Interference + Noise Ratio SINR =
P
Pv d(sv ,rv )α Pu u∈S d(su ,rv )α +N
A link set S is feasible if SINR ≥ β for all links in S. Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
5 / 20
The Scheduling Problem Given a link set L = {`1 , `2 . . . `n }, partition L in to a minimum number of feasible sub-sets. Equivalently, find the minimum number of slots required to schedule each link successfully.
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
6 / 20
The Scheduling Problem Given a link set L = {`1 , `2 . . . `n }, partition L in to a minimum number of feasible sub-sets. Equivalently, find the minimum number of slots required to schedule each link successfully. Assumptions: Powers are given: “fixed power” Powers are oblivious, ie, Pv is a function of d(sv , rv ). Also, the powers are 1 2
Length monotone. Sub-linear.
Pv = constant (uniform power) Pv = p d(sv , rv )α (linear power) Pv = d(sv , rv )α (mean power)
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
6 / 20
The Scheduling Problem (contd.) State of the art: 1 2
Centralized: O(log n) approximation (Halld´ orsson and Mitra, 2011) Distributed: O(log2 n) approximation. (Kesselheim and V¨ocking, 2010)
“Distributed”: links know their power assignments, an estimate of n and have synchronized time slots.
Thm. Let the optimal solution be T . There is randomized distributed algorithm that schedules the links in O(T × log n) slots, with high probability. In general, no distributed algorithm can do better.
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
7 / 20
Algorithm for a link (KV10)
Choose probability q = “the right value” 2: for lnqn slots do 3: transmit with i.i.d. probability q 4: end for 1:
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
8 / 20
Signal to Interference + Noise Ratio SINR =
P
Pv d(sv ,rv )α Pu u∈S d(su ,rv )α
S is feasible if SINR ≥ β for all links in S.
Affectance 1 = SINR
Pu u∈S d(su ,rv )α Pv d(sv ,rv )α
P
SINR ≥ β ≡ au (v ) = β `v succeeds iff
P
u∈S
Halld´ orsson and Mitra ()
X Pu · d(sv , rv )α β ≤1 Pv · d(su , rv )α u
Pu · d(sv , rv )α Pv · d(su , rv )α
(1) (2)
au (v ) ≤ 1
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
9 / 20
Signal to Interference + Noise Ratio SINR =
P
Pv d(sv ,rv )α Pu u∈S d(su ,rv )α
S is feasible if SINR ≥ β for all links in S.
Affectance au (v ) = β
Pu · d(sv , rv )α Pv · d(su , rv )α
Define P Incoming affectance aL (v ) = Pu∈L au (v ) Outgoing affectance av (L) = u∈L av (u)
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
9 / 20
Analysis (KV10) ¯= Let, average affectance, A
1 n
P
v ∈L aL (v ).
¯ Call Now, there must be at least n/2 links `v for which aL (v ) ≤ 2A. ¯ these links L Set q =
1 ¯. 4A
¯ E(aS (v )) = qaL (v ) ≤ Then for `v ∈ L, of active links.
1 ¯ ¯ 2A 4A
= 21 , where S is the set
Upshot In expectation, Ω( A1¯ ) fraction of the links succeed in the first slot. Thus, ¯ log n slots. we need A
Question ¯ relate to the optimal scheduling length T ? How does A
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
10 / 20
Analysis (KV10) – contd. Question ¯ relate to the optimal scheduling length T ? How does A
Answer ¯ = O(T log n) →→ log2 n approximation A
Proof Pattern 1 Find measure M, show that with transmission probability Θ( M ), algorithm takes O(aM) steps.
Show that M = O(bT ). Derive O(a × b) approximation factor.
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
11 / 20
Our Analysis Proof Pattern 1 Find measure M, show that with transmission probability Θ( M ), algorithm takes O(aM) steps.
Show that M = O(bT ). Derive O(a × b) approximation factor.
New measure Λ = arg mint |{`u ∈ L : aL (u) ≤ 4t}| ≥ |L|/4 . We show: a = log n, b = O(1).
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
12 / 20
The crux
¯ = O(T log n). A Λ = O(T ).
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
13 / 20
The crux Lemma (KV10) Assume S is feasible, and `u ≤ S. Then, aS (u) = O(1) and au (S) = O(log n).
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
14 / 20
The crux Lemma (KV10) Assume S is feasible, and `u ≤ S. Then, aS (u) = O(1) and au (S) = O(log n).
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
14 / 20
The crux Lemma (KV10) Assume S is feasible, and `u ≤ S. Then, aS (u) = O(1) and au (S) = O(log n).
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
14 / 20
The solution Λ = O(T ).
Lemma Assume S is feasible. Then there exists S 0 ≥ S/2 such that au (S) ≤ 4 (we call S 0 anti-feasible).
Lemma Assume S is feasible and anti-feasible and `u ≤ S. Then, aS (`) = O(1) and au (S) = O(1).
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
15 / 20
The solution Only feasible = bad. Feasible+Anti-feasible = good. Every feasible set has a large subset that is Feasible+Anti-feasible. Λ = arg mint |{`u ∈ L : aL (u) ≤ 4t}| ≥ |L|/4 .
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
16 / 20
The full algorithm Algorithm 1 Distributed 1: k ← 0 2: while transmission not successful do 3: q = 21k 4: for lnqn slots do 5: transmit with i.i.d. probability q 6: end for 7: k ←k +1 8: end while
Issues Scaling down to the right probabiliy Acknowledgements
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
17 / 20
Lower bound
Point: At any time t, the probability of only one links in a pair transmitting is ≤ 1/2 Thus, in expectation, no more than 1/2 of the pairs will succeed in each time slot. Hence we will need log n slots in expectation. Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
18 / 20
Discussion
Can we get real multiplicative lower bounds? Use carrier sense for better bounds? Utilize and transcend anti-feasibility.
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
19 / 20
Thanks!
Halld´ orsson and Mitra ()
Nearly Optimal Bounds for Distributed Wireless Scheduling in the SINR Model
20 / 20
