random matrices, phase transitions & queuing theory raj
TRANSCRIPT
![Page 1: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/1.jpg)
Random matrices, phase transitions & queuing theory
Raj Rao Nadakuditi
Dept. of Electrical Engg. & Computer Science
http://www.eecs.umich.edu/~rajnrao
Joint works with Jinho Baik, Igor Markov and Mingyan Liu
1
![Page 2: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/2.jpg)
Queuing theory
2
![Page 3: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/3.jpg)
A fundamental problem in queuing theory
F I F O
Stage 1 Time 0
…
Time 0 …
Time 0 …
Stage 2
Stage 3
End-to-end latency
Job 1 Job 2 Job n
Job 1 Job 2 Job n
Job 1 Job 2 Job n
-idle
2 1 … m F I F O F I F O
• m = # servers, n = # customers (or jobs)
• Objective: Characterize L(m,n) = exit time for n-th customer from m-th queue
– Model for production systems, multi-hop networks, pipelined computation
3
![Page 4: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/4.jpg)
Why is characterizing latency important?
F I F O
Stage 1 Time 0
…
Time 0 …
Time 0 …
Stage 2
Stage 3
End-to-end latency
Job 1 Job 2 Job n
Job 1 Job 2 Job n
Job 1 Job 2 Job n
-idle
2 1 … m F I F O F I F O
• Many existing applications are delay-sensitve
– Production systems, Streaming audio and video - particularly audio
– ⇒ Optimal scheduling/provisioning⇔ delay-throughput tradeoff
• Emerging applications envision control and inference over large networks
– Telemedicine, sensor networks and distributed computation
– ⇒ Quality of Service (QoS) guarantees important
• Network topology design
– Ad-hoc, multi-hop networks prevalent (e.g. deliver interet to rural areas)
– Optimal placement of hops? Remote diagnosis of service bottlenecks?
– ⇒ Statistical characterization of delay important
4
![Page 5: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/5.jpg)
A basic model
F I F O
Stage 1 Time 0
…
Time 0 …
Time 0 …
Stage 2
Stage 3
End-to-end latency
Job 1 Job 2 Job n
Job 1 Job 2 Job n
Job 1 Job 2 Job n
-idle
2 1 … m F I F O F I F O
Notation:
• Si = Server i ∈ {1, . . . ,m}• Cj = Customer j ∈ {1, . . . , n}• w(i, j) = Service time for Cj at Si
Assumptions:
• Infinitely long buffer
• Arrival process is Poissonian with rate α
• w(i, j)ind.∼ exp(1/µi)⇔ M/M/m queue
Question: Average Delay?
5
![Page 6: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/6.jpg)
Little’s Law and average delay
Informally:
Avg. Time in System =Avg. # of Cust.
Eff. Arrival Rate
By Burke’s Theorem:
P(#Cust. in Queue i = k) =
(1−
µi
α
)k(µiα
)for k = 0, 1, . . .
Consequently:
⇒ Avg.# Cust. in System =
m∑i=1
µi
µi − αi
6
![Page 7: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/7.jpg)
What Little’s law says and does not say
Avg.# Time in System =1
α
m∑i=1
µi
µi − αi
Mathematically:
Avg.# Time in System = limt→∞
∑α(t)i=0 Time spent by Customer i
α(t)
• α(t) = # Customers who arrived in the interval [0, t]
• No insights on: variance, pdf, bottleneck behavior, etc.
• Contrast with L(m,n) = exit time for Customer n from Server m
– Transient-like statistic! Computable?
7
![Page 8: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/8.jpg)
What Little’s law says and does not say
Avg.# Time in System =1
α
m∑i=1
µi
µi − αi
Mathematically:
Avg.# Time in System = limt→∞
∑α(t)i=0 Time spent by Customer i
α(t)
• α(t) = # Customers who arrived in the interval [0, t]
• No insights on: variance, pdf, bottleneck behavior, etc.
• Contrast with L(m,n) = exit time for Customer n from Server m
– Transient-like statistic! Computable?
– Yes! - Using random matrix theory!
8
![Page 9: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/9.jpg)
What makes the problem difficult?
F I F O
Stage 1 Time 0
…
Time 0 …
Time 0 …
Stage 2
Stage 3
End-to-end latency
Job 1 Job 2 Job n
Job 1 Job 2 Job n
Job 1 Job 2 Job n
-idle
2 1 … m F I F O F I F O
• m = # servers, n = # customers (or jobs)
• Objective: Characterize L(m,n) = exit time for n-th customer from m-th queue
– Strong interaction between arrival and departure process⇒ no independence
9
![Page 10: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/10.jpg)
Main message
New insights beyond Little’s Law:
• Latency mean and variance can be explicitly computed! X
• Analysis reveals emergence of phase transitions X
• Rigorous basis for statistical anomaly testing X
• Can show that O(n1/3) jobs have statistically independent latencies X
• Extends easily to quasi-reversible networks (thanks Demos!) X
• Analysis of queue-state dependent servicing (inspired by backpressure algorithms) X
• Results appear to hold even for non-exponential service times X
– Universality conjecture!
All made possible due to connection with random matrix theory!
10
![Page 11: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/11.jpg)
Phase transitions
11
![Page 12: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/12.jpg)
A numerical example
• G = Gaussian random matrix
– G = randn(n,n) or G = sign(randn(n,n))
• Xn =G+G′√
2n
• Xn = Xn + Pn– Pn = θ u u′
– u is a fixed, non-random unit norm vector
– Xn has i.i.d. zero mean, variance 1/2n entries (on off-diagonal)
Question: Largest eigenvalue? Variation with θ?
12
![Page 13: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/13.jpg)
One experimental realization
−3 −2 −1 0 1 2 3 4 5 0
0.05
0.1
0.15
0.2
0.25
0.3
ρ
1
• θ = 4, n = 500
• Bulk obeys semi-circle law on [−2, 2]
• Largest eig. ≈ 4.2
13
![Page 14: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/14.jpg)
An eigenvalue phase transition
0 1 2 3 4 5 6 7 8 9 100
2
4
6
8
10
12
TheoryExperiment: n = 50Experiment: n = 500
λ1(X
n+Pn)
θ
1
• Clear phase transition @ θ = 1 with increasing n
14
![Page 15: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/15.jpg)
Phase transition prediction
Theorem: Consider Xn = Xn + θuu′
λ1a.s.−→
θ +1
θ, θ > 1
2, otherwise
|〈u1, u〉|2a.s.−→
(
1−1
θ2
), θ > 1
0, otherwise
• Eigenvalue result first due to Peche (2006), Peche-Feral (2007)
• Eigenvector result new (and derived by us)
• Eigenvalues and eigenvectors are biased
15
![Page 16: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/16.jpg)
Phase transitions & Random matrix theory
or
What theory predicts the phase transition?
16
![Page 17: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/17.jpg)
Definitions and assumptions
Spectral measure: Eigenvalues of Xn are λ1, . . . , λn:
µXn =1
n
n∑i=1
δλi
Assumptions:
1. µXna.s.−→ µX
2. suppµX compactly supported on [a, b]
3. max(eig)a.s.−→ to b
17
![Page 18: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/18.jpg)
A basic signal-plus-noise model
Xn = θuuH +Xn
Assumptions:
• Xn is symmetric with n real eigenvalues
• θ1 > . . . > θk > 0
• Xn = QΛQ′ where Q is a Haar distributed unitary (or orthogonal) matrix
• u is a unit-norm vector
• Xn = GG∗ will satisfy conditions
18
![Page 19: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/19.jpg)
Phase transition in the eigenvalues
Theorem [Benaych-Georges and N.]: As n −→ ∞,
λ1(Xn)a.s.−→
{G−1µ (1/θi) if θ > θc := 1/Gµ(b
+),
b otherwise,
• Critical threshold depends explicitly on spectral measure of “noise”
Cauchy transform of µ:
Gµ(z) =
∫1
z − ydµ(y) for z /∈ suppµX.
19
![Page 20: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/20.jpg)
Phase transition of eigenvectors
Theorem [Benaych-Georges and N.]: As n −→ ∞, for θ > θc:
|〈u1, u〉|2a.s.−→ −
1
θ2iG′µ(ρ)
• ρ = G−1µ (1/θi) is the corresponding eigenvalue limit
Theorem: As n −→ ∞, for θ ≤ θc:
〈u1, u〉a.s.−→ 0
• Eigenvalue density at edge needed of form (x− b)α with α ∈ (0, 1]
20
![Page 21: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/21.jpg)
Above phase transition
ρa b
1
(a) Eigenvalue: θ > θc
u
u
1
(b) Eigenvector: θ > θc
21
![Page 22: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/22.jpg)
Below phase transition
a b
1
(c) Eigenvalue: θ ≤ θc
u
u
1
(d) Eigenvector: θ ≤ θc
22
![Page 23: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/23.jpg)
The queuing theory connection
23
![Page 24: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/24.jpg)
Problem setup
F I F O
Stage 1 Time 0
…
Time 0 …
Time 0 …
Stage 2
Stage 3
End-to-end latency
Job 1 Job 2 Job n
Job 1 Job 2 Job n
Job 1 Job 2 Job n
-idle
2 1 … m F I F O F I F O
Assumptions:
• Infinitely long buffer
• Arrival process is Poissonian with rate α
• w(i, j)ind.∼ exp(1/µi)⇔ M/M/m queue
Objective: Compute L(m,n) = exit time for batch of n customers when
• Queues are in equilibrium before the batch of n customers arrive
24
![Page 25: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/25.jpg)
The random matrix connection
Theorem [Baik & N., 2012]:
L(m,n)D= λ1(W )
• W = Γ1/2gg∗Γ1/2 + Σ1/2GG∗Σ1/2
• G is an m× (n− 1) matrix of i.i.d. CN(0, 1) entries
• g is an m× 1 vector of i.i.d CN(0, 1) entries
• Σ = diag(1/µ1, . . . , 1/µm)
• Γ = diag (1/(µ1 − α), . . . , 1/(µm − α))
• Sanity check: α = 0, n = 1, L(m,n) =∑
i 1/µi|gi|2
– |gi|2 is chi-squared with 2 d.o.f. ⇔ Exponential!
25
![Page 26: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/26.jpg)
The random matrix connection
Theorem [Baik & N., 2012]:
L(m,n)D= λ1(W )
• W = Γ1/2gg∗Γ1/2 + Σ1/2GG∗Σ1/2
– Rank-one-signal plus noise⇒ expect phase transition!
• G is an m× (n− 1) matrix of i.i.d. CN(0, 1) entries
• g is an m× 1 vector of i.i.d CN(0, 1) entries
• Σ = diag(1/µ1, . . . , 1/µm)
• Γ = diag (1/(µ1 − α), . . . , 1/(µm − α))
26
![Page 27: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/27.jpg)
New insight: phase transitions in queuing behaviorRecall:
• Arrival process is Poissonian with rate α < µi
• w(i, j)ind.∼ exp(1/µi)⇔ M/M/m queue
A critical rate:
lcrit = z such that∑i
1
(µi − z)2−n
z2= 0, z = lcrit ∈ (0, µmin)
Theorem [Baik & N. , 2012]:
• Case 1: 0 < lcrit < α⇔ arrival rate is faster than critical rate
– L(m,n) is normally distributed: mean O(n), variance O(m)
• Case 2: lcrit > µmin ⇔ slowest server is slower than critical rate
– L(m,n) is normally distributed: mean O(n), variance O(m)
• Case 3: α < lcrit < µmin ⇔ slowest server fast enough, arrival rate slow enough
– L(m,n) is Tracy-Widom distributed: mean O(n) and variance O(m2/3)
27
![Page 28: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/28.jpg)
Tracy-Widom versus Normal
−5 0 50
0.1
0.2
0.3
0.4
0.5
PD
F
Normal distr.
TW2 distr.
28
![Page 29: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/29.jpg)
New insights: phase transitions and more
A critical rate:
lcrit = z such that∑i
1
(µi − z)2−n
z2= 0, z = lcrit ∈ (0, µmin)
Theorem [Baik & N. , 2012]:
• Case 1: 0 < lcrit < α⇔ arrival rate is faster than critical rate
– L(m,n) is normally distributed: mean O(n), variance O(m)
• Case 2: lcrit > µmin ⇔ slowest server is slower than critical rate
– L(m,n) is normally distributed: mean O(n), variance O(m)
• Case 3: α < lcrit < µmin ⇔ slowest server fast enough, arrival rate slow enough
– L(m,n) is Tracy-Widom distributed: mean O(n) and variance O(m2/3)
29
![Page 30: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/30.jpg)
The importance of the variance scaling result
An elementary bound:
var maxXi ≤∑i
varXi
Upper-bounding latency:
varL(m,n) ≤ O(n)
• Insight 1: Upper bound matched only when there is a bottleneck!
• Insight 2: Realized variance is much less than upper bound!
– ⇒ Service prov. due to upp. bound very conservative
– Opportunity for perf. gains or relax system specs to meet existing QoS reqs!
∗ Work with Mingyan Liu on optimal file-split.in multi-route, multi-hop ntwk
30
![Page 31: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/31.jpg)
Numerical results
Mean Variancem n Experiment Theory Experiment Theory
5 5 13.1024 12.3685 9.4351 15.098110 10 30.9954 30.3849 18.6033 23.966820 20 68.3172 67.8858 33.0268 38.044940 40 145.0274 144.7371 55.1251 60.392680 80 300.9902 300.7699 90.0644 95.8673
160 160 615.9515 615.7717 148.8302 152.1799320 320 1249.4124 1249.4742 236.0294 241.5705480 480 1885.7545 1885.0567 311.7331 316.5469640 640 2521.6221 2521.5399 374.6064 383.4693
1000 1000 3955.4348 3955.3710 506.5496 516.3498
Empirical mean and variance of compared to theoretical predictions.
• Here µ1 = . . . = µm = 1
• “8 =∞”
31
![Page 32: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/32.jpg)
Numerical results
1 2 3 4 53.5
4
4.5
5
5.5
6
6.5
λ
Mea
n en
d−to
−en
d la
tenc
y pe
r st
age
Experiment: n = m = 80Experiment: n = m = 1000Theory
• Here n = m, µ1 = . . . = µm−1 = 1 , µm = 1/λ; exponential service time
• Regime where the bottleneck does not affect distribution!
32
![Page 33: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/33.jpg)
Numerical results
1 2 3 4 53.5
4
4.5
5
5.5
6
6.5
λ
Mea
n en
d−to
−en
d la
tenc
y pe
r st
age
Lognormal expt.: m = 80Lognormal expt.: m = 320Exponential theory
Predicted stoch. phase transition @ λ = 2
• Here n = m, µ1 = . . . = µm−1 = 1 , µm = 1/λ; lognormal service time
• Conjecture: Distribution-independent limiting distribution
33
![Page 34: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/34.jpg)
A fundamental recursion
Notation:
• Si = Server i ∈ {1, . . . ,m}• Cj = Customer j ∈ {1, . . . , n}• w(i, j) = Service time for Cj at Si
• L(i, j) = Exit time for Cj from Si
Fact (Glynn & Whitt, Tembe & Wolff):
L(i, j) = w(i, j) +
{L(i− 1, j) when L(i, j − 1) < L(i− 1, j),
L(i, j − 1) when L(i, j − 1) > L(i− 1, j).
Equivalently,
L(i, j) = max{L(i− 1, j), L(i, j − 1)}+ w(i, j)
34
![Page 35: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/35.jpg)
The directed last-passage percolation problem
…
…
…
…
…
…
…
…
…
… …
…
…
…
…
…
…
…
…
…
…
(m,1)
(1,n) (m,n)
2 3 4
2
3
4
(1,1)
L(m,n) = max{L(m− 1, n), L(m,n− 1)}+ w(m,n)
35
![Page 36: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/36.jpg)
The directed last-passage percolation problem
…
…
…
…
…
…
…
…
…
… …
…
…
…
…
…
…
…
…
…
…
(m,1)
(1,n) (m,n)
2 3 4
2
3
4
(1,1)
L(m,n) = maxπ∈P (m,m)
( ∑(k,`)∈π
w(k, `)
)
• P (m,n) is the set of ‘up/right paths’ ending at (m,n)
36
![Page 37: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/37.jpg)
The random matrix connection
L(m, j) = maxπ∈P (m,n)
( ∑(k,`)∈π
w(k, `)
)
Theorem [Borodin & Peche]: Assume
• w(i, j) ∼ exp(1/(ai + bj))
• Xij ∼ CN(
0, 1ai+bj
)
⇒ L(m,n)D= λ1(XX
∗)
• Related work by Johansson (2000)
• Result easily extended to Poissonian (discrete) random variables
37
![Page 38: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/38.jpg)
The percolation mapping for our problem
S 1 S 2 S 3 S 4 S 5 S 6 S 7
C 1
C 2
C 3
C 4
C 5
C 6 µ1
µ1
µ1
µ1
µ1
µ2
µ2
µ2
µ2
µ2
µ3
µ3
µ3
µ3
µ3
µ4
µ4
µ4
µ4
µ4
µ5
µ5
µ5
µ5
µ5
µ6
µ6
µ6
µ6
µ6
µ7
µ7
µ7
µ7
µ7
µ1 µ2 µ3 µ4 µ5 µ6 µ7α− α− α− α− α−α−α−
• Note that queues are in equilibrium before first customer enters
• Queue lengths are random and have (shifted) geometric distribution
• ⇒ First customer served at Si with rate µi − α, rest with µi– PASTA property = Poissonian Arrivals See Time Averages
38
![Page 39: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/39.jpg)
Ergo the random matrix connection
Theorem [Baik & N., 2012]:
L(m,n)D= λ1(W )
• W = Γ1/2gg∗Γ1/2 + Σ1/2GG∗Σ1/2
– Rank-one-signal plus noise⇒ expect phase transition!
• G is an m× (n− 1) matrix of i.i.d. CN(0, 1) entries
• g is an m× 1 vector of i.i.d CN(0, 1) entries
• Σ = diag(1/µ1, . . . , 1/µm)
• Γ = diag (1/(µ1 − α), . . . , 1/(µm − α))
39
![Page 40: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/40.jpg)
Why the random matrix connection?
40
![Page 41: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/41.jpg)
Where are the non-intersecting random walks?F I F O
Stage 1 Time 0
…
Time 0 …
Time 0 …
Stage 2
Stage 3
End-to-end latency
Job 1 Job 2 Job n
Job 1 Job 2 Job n
Job 1 Job 2 Job n
-idle
2 1 … m F I F O F I F O
• FIFO protocol means exit time trajectories do not intersect
• Mathematics of random walks⇔ classical probability theory
• Mathematics of random walks conditioned not to intersect⇔ random matrix theory
41
![Page 42: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/42.jpg)
Bijection with TASEP & corner growth model
http://www-wt.iam.uni-bonn.de/~ferrari/animations/ContinuousTASEP.html
42
![Page 43: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/43.jpg)
Non-interesecting random walks are everywhere!
• Taken from Andrei Okounkov’s 2006 Fields Medal Citation
43
![Page 44: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/44.jpg)
The traveling salesman problem
…
…
…
…
…
…
…
…
…
… …
…
…
…
…
…
…
…
…
…
…
(m,1)
(1,n) (m,n)
2 3 4
2
3
4
(1,1)
L(m,n) = maxπ∈P (m,n)
( ∑(k,`)∈π
w(k, `)
)
• Fix w(k, l), what order of processing minimizes delay?
– Limits of scheduling? Application-motivated extensions of RMT!
44
![Page 45: Random matrices, phase transitions & queuing theory Raj](https://reader031.vdocuments.us/reader031/viewer/2022022112/62121d7ab0536b623074a9a9/html5/thumbnails/45.jpg)
Main message
New insights beyond Little’s Law:
• Latency mean and variance can be explicitly computed! X
• Analysis reveals emergence of phase transitions X
• Rigorous basis for statistical anomaly testing X
• Can show that O(n1/3) jobs have statistically independent latencies X
• Extends easily to quasi-reversible networks (thanks Demos!) X
• Analysis of queue-state dependent servicing (inspired by backpressure algorithms) X
• Results appear to hold even for non-exponential service times X
– Universality conjecture!
All made possible due to connection with random matrix theory!
45