Session 8B: Performance Measurement and Management SIGMETRICS’19 Abstracts, June 24–28, 2019, Phoenix, AZ, USA

Proactive Caching for Low Access-Delay Services under

Uncertain Predictions
Ran Liu Edmund Yeh Atilla Eryilmaz
Electrical and Computer Engineering Electrical and Computer Engineering Electrical and Computer Engineering
Northeastern University Northeastern University Ohio State University
Boston, MA, USA Boston, MA, USA Columbus, OH, USA
[email protected] [email protected] [email protected]

ABSTRACT queueing theory perspective. In contrast to [2–4], we not only

Network traffic for delay-sensitive services has become a domi- examine the basic queueing dynamics of the proactive system but
nant part in the network. Proactive caching with the aid of pre- also explore how to strategically utilize uncertain predictions to
dictive information has been proposed as a promising method to enhance delay performance. In terms of delay performance, we will
enhance delay performance. In this paper, we analytically investi- show improvements relative to the Earliest-Deadline-First (EDF)
gate the problem of how to efficiently utilize uncertain predictive type strategy, which has been widely used in network scheduling
information to design proactive caching strategies with provably problems.
good access-delay characteristics. We first derive an upper bound
for the average amount of proactive service per request that the 2 NETWORK MODEL
system can support. We then analyze the behavior of a family of We consider a single-queue system with one server providing delay-
threshold-based proactive strategies using an innovative Markov sensitive services to the user. The system operates in continuous
chain analysis, and show that the average amount of proactive time from time 0. The user receives service from the server at a
service per request can be maximized by properly selecting the constant rate of µ bits/sec.
threshold. Finally, we propose the UNIFORM strategy, which is the Request Processes: Requests arrive at the server, requesting
threshold-based strategy with the optimal threshold. Surprisingly, same-sized data objects of s bits. The Potential Request Process is
we show that it outperforms the commonly used Earliest-Deadline- a Poisson Process {P (t) ; t > 0} with an overall arrival rate of λ,
First (EDF) type proactive strategies in terms of delay. We perform where the ith arrival, i.e. Potential Request i, requests object r i ∈
extensive numerical experiments to demonstrate the influence of Z+ = {1, 2, 3, . . .} at time ti ∈ R+ , where 0 < t 1 < t 2 < . . .. The
thresholds on delay performance, and explicitly compare perfor- Actual Request Process {A (t) ; t > 0} is a thinned version of P (t)
mance of the EDF strategy and the UNIFORM strategy. where each arrival on P (t) is an arrival on A (t) with probability
p, independent of all other arrivals. Let {Ri ; i = 1, 2, . . .} be IID
KEYWORDS Bernoulli (p) indicator random variables where Ri = 1 if the ith
Proactive Caching; Prefetching; Queueing Theory; Markov Chain arrival on P (t) is an arrival on A (t). We assume that every potential
request requests a different object, i.e. r i , r j , ∀i , j, i.e., the catalog
ACM Reference Format:
Ran Liu, Edmund Yeh, and Atilla Eryilmaz. 2019. Proactive Caching for Low size is assumed to be infinite.
Access-Delay Services under Uncertain Predictions. In ACM SIGMETRICS / Predictions: At time 0, the server is assumed to know the se-
International Conference on Measurement and Modeling of Computer Systems quence of objects (r 1 , r 2 , r 3 , . . . .) to be requested by the arrivals
(SIGMETRICS ’19 Abstracts), June 24–28, 2019, Phoenix, AZ, USA. ACM, New in {P (t)}, and the probability p. It has no prior knowledge of the
York, NY, USA, 2 pages. https://doi.org/10.1145/3309697.3331471 precise arrival epochs {ti }, or the realizations of the indicator ran-
dom variables {Ri }. The server observes A (t) but not P (t). At time
1 BACKGROUND AND MOTIVATION t > 0, the sequence of indices for future potential requests from
Proactive caching techniques are seen as an effective method to the server’s viewpoint, or the prediction window 1 , is:
improve the delay performance of delay-sensitive services in com- Π (t) = (I (t) + 1, I (t) + 2, I (t) + 3, . . .) (1)
munication networks. Specifically, proactive caching techniques
take advantage of predictive information of user requests and net- where I (t) is defined as:
work states, utilize spare bandwidth resources and potentially place
I (t) ≜ max {i |ti < t, Ri = 1} (2)
data objects in caches before requests are generated.
Our work aims to study the characteristics of proactive caching i.e. the index of the most recent actual request before time t. The
based on uncertain predictive information from a fundamental server proactively works on request i only if i ∈ Π (t) at time t.
More details regarding the assumptions on the network model and
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed predictions model can be found in [1].
for profit or commercial advantage and that copies bear this notice and the full citation Proactive Strategies: In a reactive scheme, the server works
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
only on requests made by actual request arrivals. We focus on
SIGMETRICS ’19 Abstracts, June 24–28, 2019, Phoenix, AZ, USA
1 Weassume the prediction window size to be infinite for simplicity of analysis. This
© 2019 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-6678-6/19/06. assumption guarantees that the server always has predicted requests on which to do
https://doi.org/10.1145/3309697.3331471 proactive work.

89
Session 8B: Performance Measurement and Management SIGMETRICS’19 Abstracts, June 24–28, 2019, Phoenix, AZ, USA

µ µ
proactive schemes, where the server works on future potential re- Corollary 1. Given µ, λ, s and p satisfying s ≤ λ < ps , the
quests when not serving requests made by actual requests. The limiting average amount of proactive work per potential request under
server can proactively send a data object, partially or in entirety, to strategy ΨP ∈ ΓP satisfies
the user, which can store the data object in a local cache. 2 µ − pλs
Let Ui (t) ≤ s be the proactive work done for request i by time U ≤ ≜ U ∗ , w.p. 1 (7)
λ (1 − p)
t, i.e. the number of bits of object r i sent to the user and stored
in the cache by time t. Let Ui ≜ min{s, Ui (t H (i) )} be the total where equality holds in (7) if and only if strategy ΨP satisfies both
Property 1 and Property 2.
proactive work done for request i, where H (i) ≜ min{j ≥ i |R j = 1}
represents the first potential request after i which is realized. Let In order to construct an optimal proactive strategy to solve (4),
U(t) ≜ (UI (t )+1 (t), UI (t )+2 (t), UI (t )+3 (t), . . .) be the set of Ui (t)’s ϕ
we first define a family of threshold-based strategies ΨP in ΓP . The
where i ∈ Π(t). V (t) is defined as the total number of bits for actual ϕ
detailed definition of ΨP can be found in [1]. Summarily speaking,
requests waiting to be transmitted in the queue at time t. At time t, ϕ
based on U(t), the prediction window Π(t) and the queue size V (t), ΨP selects a threshold ϕ ∈ (0, s], and proactively works on request
a stationary proactive rate allocation strategy ΨP at the server is J (t) = min{i ∈ Π(t)|Ui (t) < ϕ} when in proactive state at t.
defined as: In order to analyze the proactive system, we construct a Markov
chain by considering the system state at specific epochs and carry
Ψ (V (t) , Π (t) , U (t)) = ρV (t) , ρ I (t )+1 (t) , ρ I (t )+2 (t) , ρ I (t )+3 (t) , . . .

out the corresponding analysis, the details of which can be found in
ϕ
where ρV (t) is the rate allocated to serve the queue of V (t), and [1]. We prove that the threshold-based strategy ΨP gives an optimal
ρ I (t )+i (t) , i ≥ 1, is the rate allocated to fetch object r I (t )+i at time solution to problem (4) if and only if ϕ = U , as in the following
∗
t. We consider the set ΓP of proactive strategies ΨP satisfying: Corollary:
Í∞
• (Reactive State) If V (t) > 0: ρV (t) = µ, i=1 ρ =0
Í∞ I (t )+i Corollary 2. The objective in (4) is maximized with a threshold-
• (Proactive State) If V (t) = 0: ρV (t) = 0, i=1 ρ I (t )+i = µ ϕ
based proactive strategy ΨP if and only if ϕ = U ∗ .
• The limiting average amount of proactive work received per
We call ΨPU the UNIFORM strategy. We then derive the closed-
∗
potential request and per actual request, respectively,:
ÍI (t ) Í form expression for limiting average delay per request under the
Ui i ∈Z+ :i ≤I (t ), R i =1 Ui
U ≜ lim i=1 , UA ≜ lim (3) UNIFORM strategy. Next, this is compared to the limiting aver-
t →∞ I (t) t →∞ A (t) age delay per request under the EDF strategy, as in the following
exist for ΨP ; Corollary:
Problem Formulation: Given λ, µ, p and s, our optimization µ µ
Corollary 3. Given µ, λ, s and p satisfying s ≤ λ < ps , the
problem then can be formulated as:
average delay DU of the UNIFORM strategy is no greater than the
maximize U (ΨP ) (4) average delay D E of the EDF strategy with probability 1:
ΨP ∈ΓP

Let Ψ∗ be an optimal solution to problem (4) and let U max ≜ U (Ψ∗ ) DU ≤ D E , w.p. 1 (8)
denote the U achieved by Ψ∗ .
We denote the average delay per actual request under Ψ∗ by D Ψ∗ . The equality holds if and only if p = 0.
We will derive the closed-form expression of D Ψ∗ , and analytically We conduct extensive simulations to show the delay performance
demonstrate its improvement relative to average delay of the EDF of the threshold-based strategies with different thresholds. The
proactive strategy. results show that proactive caching not only greatly improves delay
performance in lightly-loaded cases as found in [4], but also, under
3 MAIN RESULT the UNIFORM strategy, works exceedingly well in heavily-loaded
We first define two important properties of proactive strategies: scenarios.

Definition 1. A proactive strategy ΨP ∈ ΓP satisfies Property 1 REFERENCES

if the following condition is satisfied: [1] Ran Liu, Edmund Yeh, and Atilla Eryilmaz. 2019. Proactive Caching for Low Access-
Í∞ Delay Services Under Uncertain Predictions. Proc. ACM Meas. Anal. Comput. Syst.
U (t)
i=I (t )+1 i 3, 1, Article 2 (March 2019), 46 pages. https://doi.org/10.1145/3322205.3311073
lim = 0, w.p. 1 (5) [2] Leela Srikar Muppirisetty, John Tadrous, Atilla Eryilmaz, and Henk Wymeersch.
t →∞ t 2015. On proactive caching with demand and channel uncertainties. In 2015 53rd
Definition 2. A proactive strategy satisfies Property 2 if the fol- Annual Allerton Conference on Communication, Control, and Computing (Allerton).
1174–1181. https://doi.org/10.1109/ALLERTON.2015.7447141
lowing condition is satisfied: [3] John Tadrous and Atilla Eryilmaz. 2016. On Optimal Proactive Caching for Mobile
Networks With Demand Uncertainties. IEEE/ACM Transactions on Networking 24,
UA = U , w.p. 1 (6) 5 (October 2016), 2715–2727. https://doi.org/10.1109/TNET.2015.2478476
[4] Shaoquan Zhang, Longbo Huang, Minghua Chen, and Xin Liu. 2017. Proactive
We now derive an upper bound of the limiting average proactive Serving Decreases User Delay Exponentially: The Light-Tailed Service Time Case.
work that can be done by a proactive strategy ΨP ∈ ΓP as: IEEE/ACM Trans. Netw. 25, 2 (April 2017), 708–723. https://doi.org/10.1109/TNET.
2016.2607840
2 Since our focus is on the effects of uncertain predictions, we assume for simplicity
that the cache size is infinite.

90