654 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 33, NO.

3, MARCH 2022

On the Analysis of Cache Invalidation

With LRU Replacement
Quan Zheng, Tao Yang , Yuanzhi Kan , Xiaobin Tan ,
Jian Yang , Senior Member, IEEE, and Xiaofeng Jiang , Member, IEEE

Abstract—Caching contents close to end-users can improve the network performance, while causing the problem of guaranteeing
consistency. Specifically, solutions are classified into validation and invalidation, the latter of which can provide strong cache
consistency strictly required in some scenarios. To date, little work on the analysis of cache invalidation has been covered. In this work,
by using conditional probability to characterize the interactive relationship between existence and validity, we develop an analytical
model that evaluates the performance (hit probability and server load) of four different invalidation schemes with LRU replacement
under arbitrary invalidation frequency distribution. The model allows us to theoretically identify some key parameters that affect our
metrics of interest and gain some common insights on parameter settings to balance the performance of cache invalidation. Compared
with other cache invalidation models, our model can achieve higher accuracy in predicting the cache hit probability. We also conduct
extensive simulations that demonstrate the achievable performance of our model.

Index Terms—LRU, cache, consistency, invalidation, hit probability, server load

1 INTRODUCTION consistency is necessary to ensure the copies obtained by

users are valid.
ACHE has been extensively employed to enhance the net-
C work performance. Benefiting from cache hits, band-
width usage over links, user-perceived delays and loads on
There are two underlying approaches for the cache con-
sistency: validation and invalidation [12]. With validation,
the caches verify the validity of their stored contents with
the origin server are reduced appreciably [1]. Information-
the origin server periodically and therefore this approach
Centric Networking (ICN) paradigm has emerged for effi-
guarantees weak consistency only. However, some applica-
cient content distribution, and various candidate architec-
tions, such as financial transactions, require strong consis-
tures including Named Data Networking (NDN) or
tency which can only be provided by invalidation.
Content Centric Networking (CCN), Publish/Subscribe
Moreover, cache invalidation can be classified into four
Internet Routing Paradigm (PSIRP), and Network of Infor-
basic schemes:
mation (NetInf) have been introduced. In NDN, every
Reactive Invalidation. When a request for a content arrives
relayed content is stored in the cache of routers, and it is
and there is a corresponding copy existing in the cache, the
used to serve future requests [2]. Ubiquitous caching is one
cache then sends an If-Modified-Since request to the server,
of the most distinctive features in this novel architecture for
which in turn will reply either with a 304 Not-modified mes-
future network.
sage if the copy is deemed up to date, or with the latest full
Owing to the introduction of caching, cache consistency
data if the copy is stale [13].
also has constantly attracted the attention of many research-
Proactive Invalidation With Removing. When a master con-
ers [3], [4], [5], [6], [7], [8], [9], [10], [11]. Caching generates
tent is updated at the server, the server notifies caches of the
numerous copies of contents distributed throughout the
change. After receiving notifications, the caches remove the
network. If, for example, a master content is updated at
stale copy [12].
the origin server, the copies of that content stored at the
Proactive Invalidation With Renewing. When a master con-
caches may become outdated. Hence, maintaining cache
tent is updated at the server, the server pushes the latest
copy to the caches which have stored the stale one [14].

Quan Zheng is with the Department of Automation, University of Science
Proactive Invalidation With Optional Renewing. Unlike the
and Technology of China, Hefei 230000, China, and also with the Institute second and third schemes, the notification sent by the server
of Advanced Technology, University of Science and Technology of China, consists of either an invalidation message or a updated copy
Hefei 230088, China. E-mail: [email protected]. in this scheme. For avoiding the waste that the updated

Tao Yang, Yuanzhi Kan, Xiaobin Tan, Jian Yang, and Xiaofeng Jiang are
with the Department of Automation, University of Science and Technology copy is never subsequently requested at the caches, usually
of China, Hefei 230000, China. E-mail: {yt2015, kan}@mail.ustc.edu.cn, only the copies of popular contents are renewed, while the
{xbtan, jianyang, jxf}@ustc.edu.cn. copies of unpopular ones are removed [15].
Manuscript received 12 July 2020; revised 8 July 2021; accepted 12 July 2021. Formally speaking, most Web applications apply valida-
Date of publication 21 July 2021; date of current version 11 Aug. 2021. tion rather than invalidation to maintain cache consistency
(Corresponding author: Xiaobin Tan.)
Recommended for acceptance by S. Pakin. due to the extra overhead on the network caused by the lat-
Digital Object Identifier no. 10.1109/TPDS.2021.3098459 ter [12]. Concretely, for reactive invalidation, each cache hit
1045-9219 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tps://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
ZHENG ET AL.: ON THE ANALYSIS OF CACHE INVALIDATION WITH LRU REPLACEMENT 655

will trigger a verification process, the number of which 2 RELATED WORK

increases sharply with network scale expanding. For proac-
Thanks to the significance of caching to the web applications,
tive invalidation, the server has to maintain per-content sta-
improving the caching performance has always been a hot
tus consisting of a list of all caches that have the copies of
topic in the networking research community for decades.
this content. As the number of contents enlarges and the
Thereinto, a considerable amount of works on replacement
copies are more widely distributed throughout the network,
policies for a single cache were produced, from classical
such status could become too enormous to maintain. There-
LRU, LFU, FIFO and RANDOM algorithms to modified
fore, how to reduce above overhead is the most imperative
LRU-K [17], LFRU [18] and many other ones. However, the
issue to be addressed when deploying the cache invalida-
requirement for line-rate operation limits the complexity of
tion. In particular, the forth scheme, proactive invalidation
the cache replacement policy. Thus most of the cache systems
with optional renewing, can reduce the content status over-
commonly apply LRU algorithm that is easy to be imple-
head and the number of invalidation messages by only
mented as well as has a complexity of Oð1Þ [1].
maintaining the status of specific contents at the server [15].
Nevertheless, how many specific contents should be main-
tained to achieve a decrease of the overhead without appre- 2.1 Evaluating the Performance of LRU Cache
ciably affecting the network performance is another issue Many analytical models for evaluating the performance of
that needs to be addressed. LRU cache also have been established in these years. The
Unfortunately, most of the current work of cache invali- first exact model was provided by King [19] to estimate the
dation is focused on the specific strategies [7], [8], [9], [10], hit probability of the cache. Flajolet et al. [20] then gave a
[16]. Little work has been proposed on establishing theoreti- simpler expression. Regrettably, the computational over-
cal models to analyze the cache invalidation quantitatively. head of [19] and [20] grows exponentially with the cache
In this paper, we propose an analytical model to evaluate capacity and the number of contents increasing. Almost at
the performance of above four invalidation schemes with the same period as [20], Dan et al. [21] developed a simple
LRU replacement. The frequency of content update can fol- approximate analytical model with the computational com-
low any distribution in our model. Additionally, in order to plexity of quadratic time. About 20 years later, Rosensweig
estimate the overhead that is extra but necessary to guaran- et al. [22] developed a novel model for general-topology
tee the strong cache consistency, we take the actively renew- cache networks by utilizing the work in [21].
ing load into the consideration of total server load. With the development of networking, cache network is of
Furthermore, our model takes great advantages on extensi- a more massive scale and an efficient model to evaluate the
bility and accuracy compared with other cache invalidation caching performance is urgently needed. A more widely
models. used model was proposed by Che et al. [23] in 2002, origi-
The major contributions of our work are: nally known as the characteristic time approximation. The
analytical model in [23] provides extremely accurate results

We model four invalidation schemes with LRU at low computational overhead under the traffic of Indepen-
replacement by using conditional probability to dent Reference Model (IRM for short), and a theoretical
characterize the interactions between existence and explanation for the success of this model is given in [24].
validity under arbitrary invalidation frequency dis- Based on the characteristic time approximation, Martina
tribution. In our analysis, we compute the hit proba- et al. [25] proposed a unified methodology to analyze the
bility and server load. performance of cache systems. The study in [25] extends the

We use our model to estimate the performance of traffic condition relied on in [23] from IRM to renewal traf-
four invalidation schemes, and identify some key fic, leading to much broader applicability.
parameters that affect their performance. Particu- The above research works provide fundamental models
larly, we theoretically give the point of the highest that can be expanded to some specific cache scenarios, such
hit probability for the proactive invalidation with as cache consistency.
optional renewing. In addition, we generalize some
principles of parameter settings to balance the hit 2.2 Cache Consistency and Analytical Models
probability and server load. The problem of cache consistency is another extensively

We perform numerical simulations to demonstrate studied topic in networking, especially with the emergence
that our model can achieve high accuracy in predict- of CCN. As we mentioned earlier, the approaches for cache
ing the hit probability and server load for individual consistency can be divided into validation and invalidation.
content requests. We also apply our model to realis- Cache validation provides only weak consistency. The
tic traffic data to estimate the invalidation effect on basic validation scenario is the Time-to-live (TTL) cache
the cache hit probability. where each content is associated with an expiration time of
The remainder of this paper is organized as follows. In a constant [3] or a variable parameter [4]. The first request
the next section, we review the related work. In Section 3, presented after the expiration time is forwarded to the
we describe the invalidation model and discuss some basic server, which successively sends a valid copy to the cache.
assumptions. Section 4 models the four invalidation Analytical models for TTL-based cache were given in [5],
schemes and gives analytical expressions respectively. Sec- [6]. The problem of placing replicas under the widely used
tion 5 evaluates the accuracy of our model and presents the TTL-based consistency scheme has been investigated [26].
simulation results. Finally, conclusions and future work are Freshness and Timestamp are also common technologies
given in Section 6. for validation in cache networks. Amadeo et al. [27] added a
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
656 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 33, NO. 3, MARCH 2022

timestamp to the data packets to facilitate the evaluation of TABLE 1

their freshness, so as to decide whether to remove them Notations
from the cache. Vural et al. [28] proposed an analytical
Notation Meaning
model that captures the trade-off between multi-hop com-
munication costs and data item freshness. C; O Cache capacity and cache occupancy size
In contrast, cache invalidation, commonly classified into L Server load
M The number of different contents
the reactive and the proactive, can provide strong consis- Num ðtÞ The number of updates for content m by time t
tency. Several invalidation strategies have been proposed Nrm ðtÞ The number of requests for content m by time t
such as Leases [7] for distributed file system, Piggyback am Request probability of content m
server invalidation [8], IR-based cache invalidation [9], Bit- ; m Request rate of total contents and content m
Sequences [10] for mobile environments, cache invalidation Pe ; Pem Mean existence probability of total contents and
strategies in wireless environments [16], etc. content m
Ph ; Phm Mean hit probability of total contents and content
There is little work on modeling the cache invalidation
m
until 2018 that Detti et al. [11] proposed an analytical model Pv ; Pvm Mean validity probability of total contents and
to evaluate the performance of LRU cache that takes into content m
account of reactive invalidation and proactive invalidation m
Pe;v The joint probability of existence and validity for
with removing. In the Detti model, the invalidation event is content m
m
considered as an independent random process. While in our Pejv The conditional probability of existence given
model, the invalidation event is associated with the existence validity for content m
m
Pvje The conditional probability of validity given
event, and these two events are regarded as a pair of interac- existence for content m
tions that are characterized by conditional probability. For Tc Cache characteristic time
computing the hit probability of proactive invalidation with Tsm ; Tum Expiration time and update time of content m
removing, both Detti model and our model can achieve a mms Expected value of the expiration time for content
very high accuracy. However, Equations (5) and (6) pre- m
sented in [11] for computing the hit probability of reactive F ðtm
s Þ CDF of the expiration time for content m
Fr ði; tÞ CDF of the content arriving time in the ith
invalidation, indeed, ignore the fact that the requests for out-
interval
dated copies stored in cache can not be regarded as a hit Y ðmÞ The number of content m’s copy stored in the
event. The hit probability of reactive invalidation obtained cache
from these two equations is identical to the hit probability Y The number of all contents’ copies stored in the
without invalidation, which has been validated in our com- cache
parison results in Section 5. Besides, the Detti model can not Mp The top Mp popular contents
be developed to compute the hit probability of proactive
invalidation with optional renewing that is able to enable a
better performance than other invalidation schemes [15]. which is constant and independent of all past
The model proposed in our work is more extensible than requests. It corresponds to the IRM widely used in
the Detti model, and can achieve a higher accuracy on pre- many studies of network cache [19], [20], [21], [23].
dicting the hit probability of reactive invalidation. More- In our work, we adopt a Zipf law to model am
over, to the best of our knowledge, we are the first to
propose an accurate analytical model to estimate the perfor-
mance of proactive invalidation with optional renewing K
am ¼ ; (1)
under a few assumptions which are commonly applied in RðmÞz
cache modeling and presented in the next section. P
where K ¼ 1= M 1
m¼1 RðmÞz is the Zipf normalization
factor, RðmÞ ¼ 1; 2; . . . ; M is the popularity rank of
3 MODEL DESCRIPTION content m, and the exponent z is a parameter ranging
3.1 Assumptions and Notations between 0.65 and 1 generally.
Using the notations described in Table 1, we make the fol-
User requests for all contents arrive according to a
lowing assumptions: Poisson process with mean arrival rate . Thus, the
requests for content m also arrive according to a

We consider a collection of M same size contents. Poisson process with mean arrival rate m ¼ am 
Without loss of generality, the size of each content is based on the above assumptions.
equal to 1 and therefore the cache capacity C and
In order to simplify the analysis, we do not consider
occupancy size O can be measured in the unit of con- the delay in data transmission. In other words, if a
tent quantity [23]. For some contents of different cache miss happens, a new content copy will be
sizes, we can divide them into constant sized chunks instantly downloaded from the server. It should be
[24], [29]. The modeling process is only related to the noted that the data transmission delay will have an
characteristic time, and has nothing to do with impact on the probability of the request hitting the
the content size [23]. We assume that the contents of valid content copy. We discuss this impact in Appen-
the size of 1 in order to simplify the calculation. dix A, which can be found on the Computer Society
1

Let fRm i gi¼1 be a sequence of request arrivals for con- Digital Library at http://doi.ieeecomputersociety.
tent m. The request probability of content m is am org/10.1109/TPDS.2021.3098459.
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
ZHENG ET AL.: ON THE ANALYSIS OF CACHE INVALIDATION WITH LRU REPLACEMENT 657

Fig. 2. Consider validity in a single interval.

Fig. 1. Requests for content m, where the dotted arrows yield invalidity
and the solid arrows yield validity. we have Tum ðiÞ  Tum ði þ 1Þ for each i ¼ 0; 1; 2; . . . with
Tum ð0Þ ¼ 0. Let fTsm ðiÞ; i ¼ 0; 1; 2; . . .g represent an i.i.d.

Since the size of the invalidation message in the sequence with a common distribution F ðtm s Þ. We shall inter-
interactions between the server and cache is tinier pret Tsm ðiÞ as the content expiration time between Tum ðiÞ and
than that of the data packet, we ignore the former Tum ði þ 1Þ, i.e., Tsm ðiÞ ¼ Tum ði þ 1Þ  Tum ðiÞ. In order to avoid
impact on the server load. trivialities, we also suppose that Tsm ðiÞ 6¼ 0, which means
multiple updates for content m cannot occur simulta-
3.2 The Characteristic Time Approximation neously. Let
LRU cache can be regarded as a stack. When a content Z 1
request arrives and gets a cache hit, the content copy in mm
s ¼ E½T m
s ðiÞ ¼ s dF ðts Þ;
tm m
(4)
cache will be reinserted into the top of the stack. Otherwise, 0
denote the mean time between successive updates and
a cache miss occurs and a new content copy will be down-
Num ðT Þ denote the number of updates for content m by time
loaded from the server and inserted into the top of the stack.
T . Then we have [30]
If the stack size reaches the limit, a content copy at the bot-
tom of the stack will be evicted (We call this event eviction). Num ðT Þ 1
The time that a content copy moves from the top to the bot- lim ¼ m: (5)
T !1 T ms
tom without any cache hits taking place and finally is
evicted from the stack is described as the cache characteris- In each interval Tsm ðiÞ, the content m’s copy hit by the
tic time Tc in the characteristic time approximation [23]. first request (the dotted arrows in Fig. 1) is stale and we call
With the consideration of Tc , a content m’s copy exists in this event invalidity. Then it is removed from the cache and
the cache at time t, if and only if the last request for content a new copy is downloaded from the server. Therefore, the
m arrives at the cache in the interval ðt  Tc ; t. By assuming subsequent requests (the solid arrows) can obtain valid cop-
that the requests for content m arrive according to a Poisson ies in this interval and we call this event validity.
process with rate m , the probability Pem that content m’s Let us consider validity in a single interval first. Without
copy is in the cache can be computed as loss of generality, the start time of the interval is supposed
mT to be 0 as shown in Fig. 2. According to the above descrip-
Pem ¼ 1  e c
: (2)
tion, if a request arriving at t yields validity, there must be at
least one request arriving in the grey area.
Let Y ðmÞ denote the number of content m’s copy stored Assuming that the request arrival for content m is a Pois-
in the cache. Obviously, it takes value 1 if a content m’s son process, the number of requests in any interval of length
copy exists in the cache and 0 otherwise. Let Y denote the t is Poisson distributed with mean m t. That is, for all s; t  0
number of all contents’ copies stored in the cache. Under
the assumption that the size of each content is equal to 1, mt ðm tÞn
the expected number of all contents’ P fNrm ðt þ sÞ  Nrm ðsÞ ¼ ng ¼ e ; (6)
hP copies stored
i Pin the n!
M M
cache is written as E½Y  ¼ E m¼1 Y ðmÞ ¼ m¼1 E
PM where Nrm ðtÞ represents the number of requests that have
½Y ðmÞ ¼ m¼1 1  Pem . With regard to a stable cache system,
occurred up to time t and Nr ð0Þ ¼ 0. So the probability that
the space of the cache is always full, and thus E½Y  ¼ C.
the request for content m arriving at t yields validityin the
Then we can get a fixed point equation as follows:
ith interval can be expressed as
X
M m
ð1  e
mT
c
Þ ¼ C: (3) Pvm ði; tÞ ¼ 1  P fNrm ðtÞ ¼ 0g ¼ 1  e t : (7)
m¼1

The value of Tc can be obtained by calculating the unique For Poisson process, the time t is uniformly distributed
root of this equation [23], [24]. over ½0; Tsm ðiÞ

t
3.3 Probability of Validity Fr ði; tÞ ¼ : (8)
Tsm ðiÞ
In this section, we give a basic model to describe whether a
content copy in cache is valid when a request for it arrives. Hence the expected value of Pvm ði; tÞ in the ith interval is
In order to eliminate the influence of the cache capacity
limit, we assume that the cache capacity is infinite and all Z Tsm ðiÞ
contents’ copies have been stored already. Pvm ðiÞ ¼ Pvm ði; tÞdFr ði; tÞ
0
As shown in Fig. 1, the server updates the content m Z Tsm ðiÞ
(9)
according to a point process fTum ðiÞ; i ¼ 0; 1; 2; . . .g where 1 m
¼ ð1  e t Þdt:
Tum ðiÞ represents the time that the update takes place. Thus, 0 Tsm ðiÞ
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
658 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 33, NO. 3, MARCH 2022

Fig. 3. Classification of the requests for content m, i ¼ 0; 1; 2; . . .

Fig. 4. If there is a request arriving in grey area, Rm

1 must yield an exis-
Now we reconsider validity with all intervals. The valid- tence and validity event.
ity probability Pvm can be expressed as the value of the
weighted accumulation of Pvm ðiÞ. On the basis of (9), we
existence but invalidity: the last request Rm 4 arrives
have before Tc but is a first request in ½Tum ði þ 1Þ;
Tum ði þ 2ÞÞ.
m
u ðT Þ
NX Thus, only the requests of type existence and validity can
Tsm ðiÞ m
Pvm ¼ lim Pv ðiÞ yield cache hits, otherwise cache misses take place. For con-
T !1 T
i¼0
(10) tent m, let us denote the probability of existence and validity
m
u ðT Þ
NX Z m
1 Ts ðiÞ m
m
in the ith interval as Pe;v ðiÞ where e represents the event exis-
¼ lim ð1  e t Þdt: tence and v represents the event validity. Then the hit proba-
T !1 T 0
i¼0
bility of content m can be expressed as

Hence, on the basis of (5) and (10), the expected value of Phm ðiÞ ¼ Pe;v
m
ðiÞ ¼ Pejv
m
ðiÞ  Pvm ðiÞ; (12)
Pvm can be expressed as
or
" m Z #
u ðT Þ
NX Tsm ðiÞ
1
Phm ðiÞ ¼ Pe;v ðiÞ ¼ Pvje ðiÞ  Pem ;
m
ð1  e t Þdt
m m
E½Pvm  ¼ E lim (13)
T !1 T 0
i¼0
m "Z m # m
where Pejv ðiÞ and Pvje
m
ðiÞ are conditional probabilities. Then
u ðT Þ
NX Ts ðiÞ
1 m t
¼ lim E ð1  e Þdt the average hit probability of content m is
T !1 T 0
i¼0
Z Z m " m #
u ðT Þ
NX
N m ðT Þ 1 ts ðiÞ m Tsm ðiÞ m
¼ lim u ð1  e t ÞdtdF ðtm
s Þ Phm ¼ E lim Ph ðiÞ ; (14)
T !1 T 0 0 T !1 T
i¼0
Z 1 Z tm
1 s m
¼ m ð1  e t ÞdtdF ðtm s Þ: and average hit probability of total contents is
ms 0 0
(11) X
M
Ph ¼ am Phm : (15)
m¼1
The validity probability and hit probability mentioned
below both refer to their respective expected values.
Now we can model the four different schemes of cache
In the model assumptions, it is assumed that the data
invalidation separately and give the hit probability as well
transmission delay is out of consideration. However, based
as server load based on (12), (13), (14) and (15).
on (7), (8), (9) and (11), the influence of transmission delay
on the validity probability can be given as presented in
4.1 Reactive Invalidation
Appendix A, available in the online supplemental material.
Considering that different characteristic time Tc and content
expiration time Tsm have different effects on the analysis of
4 CACHE INVALIDATION WITH LRU REPLACEMENT m
Pejv m
and Pvje , we model the cache behavior in two cases as
In this section, we give a complete model with the consider- follows:
ation of both existence and validity to derive the cache hit
probability and server load. 4.1.1 Tc > Tsm ðiÞ
As illustrated in Fig. 3, the requests for content m are
Assuming that a request Rm 1 arrives at time t as shown in
classified into four types:
Fig. 4a. If it yields a validity event, there must be at least one

non-existence and invalidity: when Rm request arriving in ½0; tÞ (the grey area). Let t0 denote the
1 arrives, there is
no copy of content m existing in the cache. Mean- time that the first request Rm 0 arrives in ½0; tÞ. Since Tc >
while, it is the first request in the interval Tsm ðiÞ and 0  t0 < t < Tsm ðiÞ, we have t0 < t < Tc . That
½Tum ðiÞ; Tum ði þ 1ÞÞ. is to say, there must be a copy of content m existing in cache

existence and validity: the next request Rm at time t and thus Pejv m
ðiÞ ¼ 1. On the basis of (9) and (12),
2 arrives
before the characteristic time Tc and is not the first we have
request in ½Tum ðiÞ; Tum ði þ 1ÞÞ. Z Tsm ðiÞ

non-existence but validity: the third request Rm 3 is also Phm ðiÞ ¼ Pvm ðiÞ ¼
1 m
ð1  e t Þdt: (16)
not the first request in ½Tum ðiÞ; Tum ði þ 1ÞÞ but arrives 0 Tsm ðiÞ
after Tc .
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
ZHENG ET AL.: ON THE ANALYSIS OF CACHE INVALIDATION WITH LRU REPLACEMENT 659

4.1.2 Tc  Tsm ðiÞ rewritten as follows:

As shown in Fig. 4b, we can separate the interval ½0; Tsm ðiÞÞ Z Z
Tc0 tm
into two periods: ½0; Tc Þ and ½Tc ; Tsm ðiÞÞ. If the arrival time t 1
¼ m
s m
ð1  e t ÞdtdF ðtm
Phm s Þ
of Rm1 is in ½0; Tc Þ, the analysis is the same as that with Tc > ms 0 0
Tsm ðiÞ. Otherwise, t is in ½Tc ; Tsm ðiÞÞ. If there is a copy of con- Z 1 nZ Tc0
1 m (20)
tent m existing in cache at time t, there must be a request þ m ð1  e t Þdt
ms Tc0
0 arriving in ½t  Tc ; tÞ. That is to say, R1 must yield a
0
Rm m
o
0  mT 0
validity event and Pvje ¼ 1. With the assumption that the
m
s  Tc Þð1  e
þ ðtm c Þ dF ðtm
s Þ;
requests arrive according to a Poisson process, the arrival
time t of Rm and the value of Tc0 can be derived by calculating the unique
1 is uniformly distributed. Therefore, the proba-
bility of 0  t < Tc is Tc =Tsm ðiÞ and the probability of Tc  root of the following equation:
t < Tsm ðiÞ is 1  Tc =Tsm ðiÞ. Then the hit probability of con-
X
M
tent m can be written as (17) based on (2), (12), (13) and (16). O¼ Phm : (21)
m¼1
Phm ðiÞ ¼ P0t < Tc  Pðe;vÞj0t
m
< Tc ðiÞ þ PTc t < Ts ðiÞ  Pðe;vÞjTc t < Tsm ðiÞ ðiÞ
m
m

Tc
¼ Pm ðiÞ  Pvj0t
m
< Tc ðiÞ
Tsm ðiÞ ðejvÞj0t < Tc 4.3 Proactive Invalidation With Renewing
T m ðiÞ  Tc m For proactive invalidation with renewing, when the server
þ s m PðvjeÞjTc t < Tsm ðiÞ ðiÞ  PejT
m
c t < Ts ðiÞ
m
Ts ðiÞ updates a content, it will push a new content copy to the
Z Tc  
1 m Tc m cache to renew rather than to remove the stale one. Hence
¼ m ð1  e t Þdt þ 1  m ð1  e Tc Þ: the invalidity event has no influence on the cache hit behav-
Ts ðiÞ 0 Ts ðiÞ
ior. In other words, as long as a content copy exists, the
(17)
request for it will yield a cache hit. According to (2), we have
Consequently, according to (5), (14), (16) and (17), we mT
have Phm ¼ 1  e c
: (22)

Z Tc Z tm
1 s m
Phm ¼ ð1  e t ÞdtdF ðtm
s Þ 4.4 Proactive Invalidation With Optional Renewing
mms 0 0
Z 1 n Z Tc Different from proactive invalidation with renewing, proac-
1 m (18)
þ ð1  e t Þdt tive invalidation with optional renewing only updates the
mmsTc 0 top Mp popular contents. For unpopular ones, it will remove
o
m Tc
þ ðtm
s  Tc Þð1  e Þ dF ðtm
s Þ:
the stale copies of them from the cache. Let us denote the hit
probabilities of popular contents and unpopular contents as
Phm pop and Phm unpop respectively. Thus, we have
4.2 Proactive Invalidation With Removing (
Pmh pop ; m  M p ;
For proactive invalidation with removing, not only eviction Ph ¼
m
m
: (23)
but also invalidity makes content copies nonexistent in the P h unpop ; otherwise:
cache, causing the cache to be not always full. With regard
to this fact, the cache occupancy size O can be represented where
by means of two bounds: (i)when cache is full, evidently the m T 00
first bound is C, (ii)otherwise, only the valid contents can Phm pop ¼ 1  e c ; (24)
stay in the cache P and the second bound can be expressed and
approximately as M m¼1 1  Pv . We take the minimum value
m
R Tc00 R tm m
of both Phm unpop ¼ 1
mm ð1  e t ÞdtdF ðtm
s
s Þ
s 0 n R 00
0
R1
( ) þ m1m T 00
Tc m t
XM 0 ð1  e Þdt (25)
s c o
O  min C; 1  Pv
m
00 m 00
 Tc
þðtm
s  Tc Þð1  e Þ dF ðtm s Þ:
m¼1
( Z 1 Z tm )
XM
1 s
m t
¼ min C; ð1  e ÞdtdF ðts Þ :
m
The value of new characteristic time Tc00 can also be
mm 0
m¼1 s 0
obtained by calculating the fixed point equation as follows:
(19)
X
M X
Mp X
M
In addition, the time that a content copy moves from the O0 ¼ Phm ¼ Phm pop þ Phm unpop ; (26)
top to the bottom and finally to the outside of the cache m¼1 m¼1 m¼Mp þ1

without any hits taking place (eviction event) cannot be char-

where the cache occupancy size O0 can be expressed as
acterized by Tc due to the invalidity event. Let Tc0 denote the
new characteristic time.  PM
Hence, by substituting Tc0 for Tc in (17), previous analyti- O0  minfC; Mp þ m¼Mp þ1 Pvm g ; Mp < C; (27)
cal expressions obtained for reactive invalidation can be C ; otherwise:
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
660 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 33, NO. 3, MARCH 2022

Fig. 5. Cache hit probability for individual contents.

4.5 Server Load 5 PERFORMANCE EVALUATION

For reactive invalidation and proactive invalidation with
The aim of this section is threefold. First, we validate the ana-
removing, the server load is equal to the miss rate of the
lytical expressions obtained previously against simulations
cache. Then we have
to evaluate the accuracy of our model, and also we compare
XM
our model with Detti model on reactive invalidation and
L¼ m ð1  Phm Þ ¼ ð1  Ph Þ: (28)
proactive invalidation with removing. Second, we assess the
m¼1
impact of the cache invalidation on the LRU cache. Third, we
For proactive invalidation with renewing, the server not compare the four different schemes of cache invalidation
only responds to the miss stream, but also actively pushes under expiration time, cache capacity as well as the number
content copies to update the stale ones in the cache. Hence of actively renewed contents respectively and wish to get
the server load can be separated into two parts: direct load some common insights on the parameter settings.
and push load. Obviously, the calculation of direct load is In order to implement the experiment, we developed a
the same as (28). With regard to push load, the server simulation platform based on Python. A user simulator, a
pushes a new copy if and only if there is a stale one existing cache simulator as well as a server simulator are created in
in the cache. For each update point, the probability of the our platform. The user simulator requests for contents
m
content m’s copy existing in the cache is 1  e Tc based on according to the Zipf model. If the requested content is in
(2) and the update rate is 1=mm the cache and not outdated, a cache hit event will be
s . Then we have
recorded. Otherwise, a cache miss will be recorded. An
XM
1 m LRU stack is installed on the cache simulator that is used to
L ¼ ð1  Ph Þ þ ð1  e Tc Þ: (29)
m m respond to the user requests and receive the invalidation
m¼1 s
messages from the server. Besides, the server simulator is in
For proactive invalidation with optional renewing, since charge of maintaining the content update information and
the server only updates the first Mp contents, the total server sending invalidation messages actively or reactively accord-
load can be written as ing to different invalidation schemes.
Except where otherwise stated, we consider a cache
X
Mp
1 m 00 capacity with C ¼ 100, where requests arrive for 5000 con-
L ¼ ð1  Ph Þ þ ð1  e Tc Þ: (30)
m
m¼1 s
m tents following a Poisson distribution with mean rate  ¼
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
ZHENG ET AL.: ON THE ANALYSIS OF CACHE INVALIDATION WITH LRU REPLACEMENT 661

Fig. 6. Comparison between Detti model and our model on reactive Fig. 7. Comparison between Detti model and our model on proactive
invalidation. invalidation with removing.

20 req=s. The request probability is modeled by a Zipf distri- Fig. 7 shows the comparison between Detti model and
bution with exponent z ¼ 0:8. In addition, for proactive our model on proactive invalidation. The curves of Detti
invalidation with optional renewing, we set Mp ¼ 20. With- model and our model are almost coincided and make a
out loss of generality, it is supposed that the higher the quite good agreement on fitting the hit ratio of the simula-
request probability of the content, the lower the content ID. tion of proactive invalidation. In Figs. 6 and 7, the mean
expiration time is all set to 10s.
5.1 Cache Hit Probability Fig. 8 shows the impact of different mean expiration time
For measuring the accuracy of our model, we denote the hit mms on the hit probability Ph . Since the invalidity event makes
probability error as no difference to the cache hit behavior in the proactive
invalidation with renewing, the curve of which can be
jPhmodel  Phsim j regarded as a baseline that is equivalent to the hit probabil-
d¼ ; (31) ity without consideration of content expiration. From Fig. 8,
Phsim
we can see that the hit probabilities of reactive invalidation
where Phmodel and Phsim are the mean hit probabilities of the and proactive invalidation with removing increase gradu-
model and simulation respectively. ally and approach the baseline while the mean expiration
Fig. 5 characterizes the hit probability of individual con- time increasing. The smaller the mean expiration time, the
tents with four schemes of cache invalidation. Only the first greater the impact of it on the hit probability. Additionally,
50 contents are shown to avoid cluttering the figure. For com- it is noticed that proactive invalidation with removing has a
prehensive assessments, we set up three distributions of higher hit probability than reactive invalidation, because
expiration time Tsm which are constant (Figs. 5a, 5d, 5g and the former leaves more space for caching by removing stale
5j), uniformly distributed over [0,20] (Figs. 5b, 5e, 5h, and 5k) content copies from the cache.
and exponentially distributed with mean rate 0:1req=s The hit probability of proactive invalidation with optional
(Figs. 5c, 5f, 5i, and 5l). In particular, Figs. 5j, 5k, and 5l show renewing in Fig. 8 is distinctive. The curve of it increases
an apparent piecewise point at ID ¼ 20 for reasons of only when mm s is less than 4s and then decreases and approaches
the top 20 contents being renewed. As revealed in above fig- the baseline from above. To explain this phenomenon, we
ures, the analytical curves match the simulation results sur- recall the definition of the average hit probability Ph first.
prisingly within the maximum error of less than 1.2 percent, According to (15), highly popular contents contribute a lot to
which illustrates the extremely high accuracy of our model. Ph . Therefore, when mm s is small, there is enough space to
Given space limitations, we only consider Tsm with a store the popular contents in the cache, making the hit proba-
exponential distribution with mean rate 1=mm s in the follow- bility of them close to 1. With the increase of mms (but still very
ing part of this paper. small), the hit probability of unpopular contents grows, thus
Fig. 6 shows the comparison between Detti model and causing the overall hit probability to be higher. When mm s is
our model on reactive invalidation. The results of the simu- further larger, unpopular contents will crowd out popular
lations of reactive invalidation and no invalidation are also ones in the cache, which lessens the hit popularity of the lat-
presented in Fig. 6. It is obvious to see that the Detti model ter and then leads to a reduction in the overall hit probability.
for reactive invalidation, in practice, is to describe the no Also be noted that, when the cache is in a critical state of satu-
invalidation where the requests for outdated contents ration or unsaturation (corresponding to the interval [4,8] in
stored in the cache are also counted for a cache hit. The error Fig. 8), the approximate Equation (27) provides a slight error
of Detti model between the simulation of no invalidation is of less than 4.28 percent.
0.34 percent, but it is 22.11 percent between the simulation Fig. 9 shows the impact of different cache capacities C on
of reactive invalidation. With this just the opposite is, the the hit probability Ph . As the cache capacity increases, the hit
error of our model between the simulation of reactive invali- probabilities of the four invalidation schemes grow. How-
dation is only 0.24 percent. ever, due to the impact of content expiration, (i) the curves of
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
662 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 33, NO. 3, MARCH 2022

Fig. 8. Hit probability versus mean expiration time, by fixing cache capac- Fig. 10. Server load versus mean expiration time, by fixing cache capac-
ity to 100. ity to 100.

invalidation with renewing. Besides, associating Fig. 10 with

Fig. 8, it is seen that the proactive invalidation with optional
renewing can achieve a great improvement on hit ratio with-
out bringing too great extra load when the server updates
the contents at a high frequency. As mm s increases, the server
updates contents at a slower rate and generates lower update
load, which brings the server load of the four schemes of
cache invalidation basically consistent.
Fig. 11 shows the impact of different cache capacities C
on the server load. The curve of the proactive invalidation
with renewing grows nearly linearly with the increase of C,
due to the server updating all contents stored in the cache.
By contrast, the other three schemes of cache invalidation
produce lower server load steadily, among which the proac-
tive invalidation with optional renewing produces slightly
Fig. 9. Hit probability versus cache capacity, by fixing mean expiration more (less than 4.37 percent in Fig. 11).
time to 10s.
5.3 The Setting of Mp in Proactive Invalidation With
reactive invalidation and proactive invalidation with remov- Optional Renewing
ing steadily flatten out and tend to merge to the validity
By comparing Figs. 9 and 11, we observe that the proactive
probability, (ii) and the curve of proactive invalidation with
invalidation with optional renewing generates almost iden-
optional remains basically invariable, while the cache capac-
tical load for the server with the reactive invalidation and
ity further increases. Since the top 20 popular contents are
proactive invalidation with removing. However, the hit
updated actively and the contribution of hit probability from
probability of the former is up to 28.77 percent higher than
them is more remarkable, the proactive invalidation with
that of the latter. Additionally, since the small cache capac-
optional renewing has a higher hit probability even than the
ity limits the amount of the cached contents, most of the
baseline when the cache capacity is small.
contents in the cache are the popular ones. Based on this
fact, proactive optional renewing is able to outperform the
5.2 Server Load proactive renewing in both hit rate and server load when
Considering that the requests for stale contents impose the cache capacity is small. Concretely speaking, (i) in case
additional traffic burdens on the server, we compare the of the small cache capacity of Fig. 9, for the proactive
impact of the four invalidation schemes on the server load optional renewing, the positive updating mechanism allows
in this section. popular contents to stay in the cache longer and the positive
Fig. 10 shows the impact of different expiration time mm s eviting mechanism makes unpopular contents stay in the
on the server load. Owing to indiscriminately updating, the cache shorter. Under the co-action of these two mechanisms,
proactive invalidation with renewing produces much more the cache with small capacity can store more popular con-
load for the server. Similarly, the proactive invalidation with tents which contribute significantly more to the total cache
optional renewing also produces the extra update traffic of hit probability. Thus, the hit probability of the proactive
the top 20 popular contents, hence resulting in apparently optional renewing is higher than that of the proactive
more load for server compared to the reactive invalidation renewing. (ii) In case of the small cache capacity of Fig. 11,
and proactive invalidation with removing when mm s is small. as soon as the content (whether it is popular or not) in the
However, the extra load of proactive invalidation with cache is updated, the server in the proactive renewing
optional renewing is much lower than that of the proactive scheme will push a new copy. However, in the proactive
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
ZHENG ET AL.: ON THE ANALYSIS OF CACHE INVALIDATION WITH LRU REPLACEMENT 663

Fig. 11. Server load versus cache capacity, by fixing mean expiration Fig. 12. Hit probability versus Mp =C for proactive invalidation with
time to 10s. optional renewing, by fixing mean expiration time to 10s.

optional renewing scheme, the pushing event only occurs

when the popular content in the cache is updated. Thus, the
server load of the proactive optional renewing is lower than
that of the proactive renewing. In the following section, we
will pay particular attention to the performance of the pro-
active invalidation with optional renewing and derive some
principles on the parameter setting.
Since the key to the proactive invalidation with optional
renewing is to update the top Mp popular contents and the
cache capacity also has an impact on the setting of Mp , we
set different Mp =C and record the corresponding hit proba-
bility and server load as shown in Figs. 12 and 13 respec-
tively. It is noticed that with Mp growing, the hit probability
increases first and then decreases, while the server load
increases and basically flattens out. Therefore, an appropri-
ate Mp =C should be selected to balance the hit probability Fig. 13. Server load versus Mp =C for proactive invalidation with optional
and server load of which the proactive invalidation with renewing, by fixing mean expiration time to 10s.
optional renewing can achieve better performance.
Furthermore, from Figs. 12 and 13, we can see an appar- systems. Figs. 12 and 13 demonstrate that: (i) the increase of
ent turning point in each curve when C  200. To figure out Mp causes both hit probability and server load to rise rap-
this point, let us recall the curve of proactive invalidation idly, and then causes the hit probability to fall slowly and
with optional renewing in Fig. 8, where when the cache is in the server load to rise slowly; (ii) the variation magnitude of
a critical state of saturation or unsaturation, the hit probabil- the hit probability becomes larger than that of the server
ity reaches maximum. Thus we can set Mp 0 to keep the load as the cache capacity growing. Further, based on the
cache in this state and obtain these turning points. above two points respectively, it is concluded that (i) limit-
According to (27), this state can be expressed approxi- ing Mp to a smaller value could achieve a lower server load,
mately as but at the cost of lower hit probability; (ii) the cost of
increasing the server load by raising the value of Mp out-
X
M
C  Mp 0 þ Pvm : (32) weighs the benefit of increasing the cache hit probability,
m¼Mp 0 þ1 and the gap between the cost and benefit becomes larger
with the increase of cache capacity. Therefore, we could
Hence, Mp 0 can be derived by calculating the positive gain some principles on the setting of Mp : (i) when the
integer root of the above equation, if and only if C > server has sufficient load capacity, choosing Mp close to Mp 0
PM m can improve the hit probability greatly, (ii) otherwise,
m¼1 v . Then we redraw two curves of proactive invali-
P
dation with optional renewing with different cache capaci- choosing a relatively small Mp can maintain the server load
ties in Figs. 12 and 13 respectively, where the new curves at a low level but still obtains a fairly high hit probability.
(solid curves) basically pass through the turning point of
each original curve. 5.4 The Performance Evaluation in Realistic Traffic
The customizable parameter setting on Mp in the proac- In order to evaluate four invalidation schemes performance,
tive invalidation with option renewing greatly enhances the we apply our model to realistic traffic data that is obtained
flexible application of this scheme. By setting Mp to different from the open APIs provided by Wikipedia.1 It is important
values, we can obtain different combinations of hit probabil-
ity and server load to meet the demands of various cache 1. Wikipedia’s open interface: https://wikimedia.org/api/rest_v1
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
664 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 33, NO. 3, MARCH 2022

Fig. 15. The hit probabilities of four invalidation schemes for individual
Fig. 14. The distribution of content expiration time counted from Wikipe- contents under the norm distribution of expiration time with m ¼ 45547
dia from January 1, 2015 to December 31, 2019, which is fitted with a and s ¼ 2630.
norm distribution whose shape parameters are m ¼ 45547 and s ¼ 2630.

to note that the APIs only offer the number of page edited
times and corresponding date, and we use the interval
between two consecutive editing events as the content expi-
ration time. In the following part, first, we analyze the real
distribution of content expiration time. Second, we compare
four invalidation schemes with real distribution of content
expiration time. Third, we reduce the mean expiration time
to reflect the effect of invalidation on cache hit probability.
Fig. 14 shows the real distribution of content expiration
time in Wikipedia. The data ranges from January 1, 2015 to
December 31, 2019. Since the smallest time interval is day in
the data, we use the ratio of the number of all page edits to
the number of edited pages for the day to approximately
represent the single page edits per day, and then we use the
ratio of 24 3600ðsÞ to the single page edited times per day
to present the expiration time of the single page. The count- Fig. 16. The hit probability of four invalidation schemes versus mean
ing result approximately follows a norm distribution with expiration time. The values of mean expiration time are 45s, 455s, 4554s
and 45547s respectively.
m ¼ 45547 and s ¼ 2630.
Similarly, to avoid cluttering the figure, only the hit prob-
abilities of the top 50 contents are marked in Fig. 15. Since
the hit probability of proactive invalidation with renewing four invalidation schemes gradually diverge, which means
is not affected by the expiration time, the curve of it can be that the impact of invalidation on hit probability is gradu-
regarded as a curve without invalidation. It is clearly to see ally strengthened. The hit probabilities of the four invalida-
that the curves of four invalidation schemes are almost coin- tion schemes are proactive optional renewing, proactive
cide. That is because the mean content expiration time is renewing, proactive removing and reactive from the largest
45547s which is large enough to eliminate the impact of to the smallest, which is also consistent with the order in
invalidation on cache hit probability. The content is evicted Fig. 8.
from the cache due to the cache capacity limitation far Fig. 17 shows the hit probability of proactive invalidation
before it reaches the expiration time. with optional renewing for individual contents under
Fig. 16 shows the impact of different mean expiration N ð45547; 26302 Þ, N ð4554; 262:92 Þ, N ð455; 26:32 Þ and N ð45; 2:62 Þ
time on the hit probability of four invalidation schemes. It in descending order of m respectively. For proactive invali-
has been seen that the large mean expiration time has dation with optional renewing, the hit probabilities of popu-
scarcely influence upon hit probability in Fig. 15. Thus, we lar contents and unpopular contents are different. However,
decrease the mean expiration time, m, to improve the influ- this difference is not significant in the curve due to the large
ence of invalidation. Besides, in order to maintain the shape mean expiration time as shown in Figs. 5a, 5b, and 5c. Only
of the norm distribution, the other parameter, s, is scaled when the mean expiration time is relatively small as shown
down. The distributions of expiration time in Fig. 16 in Fig. 5d, the difference become obvious and an apparent
are N ð45547; 26302 Þ, N ð4554; 262:92 Þ, N ð455; 26:32 Þ and piecewise point appears at ID = 20. The advantages of pro-
N ð45; 2:62 Þ in descending order of m. As shown in Fig. 16, active invalidation with optional renewing over other three
with the reduction of mean expiration time, the curve of schemes also become notable as shown in Fig. 16 at m ¼ 45.
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
ZHENG ET AL.: ON THE ANALYSIS OF CACHE INVALIDATION WITH LRU REPLACEMENT 665

bandwidth is abundant, applying proactive invalida-

tion with optional renewing is a good choice. Other-
wise, this scheme can also be used, because the
server load can be reduced by decreasing the num-
ber of actively renewed contents at the expense of a
little hit probability.

6 CONCLUSION
In this paper, we address the problem of modeling four
schemes of cache invalidation with LRU replacement. By
applying conditional probability to characterize the interac-
tions between existence and validity, we derive analytical
expressions for cache hit probability and server load under
arbitrary frequency distribution. The simulation results
demonstrate that our model is able to achieve more extensi-
bility and exceedingly high accuracy than the Detti model.
Also, we compare the performance of four different invali-
Fig. 17. The hit probability of proactive invalidation with optional renew-
dation schemes, among which the proactive invalidation
ing for individual contents under the norm distribution of expiration time. with optional renewing has better scalability and, further,
Only the top 20 contents are positively renewed. can obtain the performance balance by reasonable parame-
ter settings.
5.5 The Summary of Simulation Results and Though the model of invalidation presented here is for
Insights LRU caches, we believe it can be adapted to other caches
In this section, we measure the performance of the proposed with which the characteristic time is able to be associated.
model, draw comparisons with an existing model, assess In addition, the insights gained through a single cache also
the invalidation impact on the LRU cache and compare four can be expanded to guide the design of invalidation
different schemes of cache invalidation. Through the above schemes in cache networks. Furthermore, how the time-
simulations, some important results and insights can be varying content popularity, future traffic (e.g., IoT) as well
summarized as follows: as the distribution of contents’ copies affect the model of
cache invalidation are also attractive topics. We would keep

Our model can achieve a fairly high accuracy on pre- on investigating these issues in the future.
dicting the hit probability and server load of four
cache invalidation schemes. ACKNOWLEDGMENTS

Our model is more accurate than the currently pro-
posed model in calculating the hit probability of the This work was supported in part by the National Key R&D
reactive invalidation. Program of China under Grant 2020YFA0711400, in part by

Due to the mechanism of positively updating popu- the Key R&D Program of Anhui Province in 2020 under Grant
lar contents, the invalidation with proactive optional 202004a05020078 and in part by the CETC Joint Advanced
renewing has a better performance against other Reasearch Foundation under Grant 6141B08080101. We also
invalidation schemes when the cache capacity is would like to acknowledge the supports from the Institute of
tight and the content expires frequently. Artificial Intelligence, Hefei Comprehensive National Science

When the cache is in the critical state of saturation Center and the Institute of Advanced Technology, University
and unsaturation, the hit probability of the invalida- of Science and Technology of China.
tion with proactive optional renewing can reach
maximum. Surprisingly, this cache state can be real- REFERENCES
ized by setting Mp in accordance with our model. [1] S. Podlipnig and L. B€ osz€ormenyi, “A survey of web cache replace-

In general, the proactive invalidation with proactive ment strategies,” ACM Comput. Surv. (CSUR), vol. 35, no. 4,
pp. 374–398, Dec. 2003.
optional renewing can be prioritized when the sys- [2] S. Lee, I. Yeom, and D. Kim, “T-caching: Enhancing feasibility of
tem requires strong cache consistency. Particularly, in-network caching in ICN,” IEEE Trans. Parallel Distrib. Syst., vol.
by setting Mp to reasonable values in different scenes 31, no. 7, pp. 1486–1498, Jul. 2020.
as we mentioned in Section 5.3, this invalidation [3] J. Jung, A. W. Berger, and H. Balakrishnan, “Modeling TTL-based
internet caches,” in Proc. IEEE 22nd Annu. Joint Conf. IEEE Comput.
scheme can achieve better performance on balancing Commun. Soc., 2003, pp. 417–426.
the hit probability and server load. [4] J. Gwertzman and M. I. Seltzer, “World wide web cache consis-

In the simulation with the real distribution of content tency,” in Proc. USENIX Annu. Techn. Conf., 1996, pp. 141–152.
[5] O. Bahat and A. M. Makowski, “Measuring consistency in TTL-
expiration time, the effect of invalidation on cache based caches,” Perform. Eval., vol. 62, no. 1–4, pp. 439–455, 2005.
performance is obvious only when the mean expira- [6] N. C. Fofack, P. Nain, G. Neglia, and D. Towsley, “Analysis of
tion time is small. TTL-based cache networks,” in Proc. 6th Int. ICST Conf. Perform.

In some networks with real-time features (e.g., IoT Eval. Methodol. Tools, 2012, pp. 1–10.
[7] C. Gray and D. Cheriton, “Leases: An efficient fault-tolerant
networks), the content expiration time is small. If the mechanism for distributed file cache consistency,” ACM SIGOPS
server performance is high and the network Operating Syst. Rev., vol. 23, no. 5, pp 202–210, 1989.
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.
666 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 33, NO. 3, MARCH 2022

[8] B. Krishnamurthy and C. E. Wills, “Piggyback server invalidation Tao Yang received the BS degree in 2019 from
for proxy cache coherency,” Comput. Netw. ISDN Syst., vol. 30, the Department of Automation, University of Sci-
no. 1–7, pp. 185–193, 1998. ence and Technology of China, Hefei, China,
[9] G. Cao, “A scalable low-latency cache invalidation strategy for where he is currently working toward the MS
mobile environments,” IEEE Trans. Knowl. Data Eng., vol. 15, degree with the Department of Automation. His
no. 5, pp. 1251–1265, 2003. current research interests include caching and
[10] J. Jing, A. Elmagarmid, A. S. Helal, and R. Alonso, “Bit-sequences: modeling of NDN.
An adaptive cache invalidation method in mobile client/server
environments,” Mobile Netw. Appl., vol. 2, no. 2, pp. 115–127, 1997.
[11] A. Detti, L. Bracciale, P. Loreti, and N. B. Melazzi, “Modeling LRU
cache with invalidation,” Comput. Netw., vol. 134, pp. 55–65, 2018.
[12] M. Rabinovich and O. Spatscheck, Web Caching and Replication.
Boston, MA, USA: Addison-Wesley, 2002. Yuanzhi Kan received the BS degree in 2015
[13] R. Fielding et al., “Hypertext transfer protocol–http/1.1,” 1999. from the Department of Automation, University of
[Online]. Available: https://www.w3.org/Protocols/rfc2616/ Science and Technology of China, Hefei, China,
rfc2616.html where he is currently working toward the MS
[14] A. Iyengar, E. Nahum, A. Shaikh, and R. Tewari, “Web caching, degree with the Department of Automation. His
consistency, and content distribution,” The Practical Handbook of current research interests include caching and
Internet Computing, Boca Raton, FL, USA: CRC Press, 2004. modeling of NDN.
[15] A. Iyengar, E. Nahum, A. Shaikh, and R. Tewari, “Enhancing web
performance,” in Proc. IFIP World Comput. Congr., 2002, pp. 95–126.
[16] K.-L. Tian, J. Cai, and B. C. Ooi, “An evaluation of cache invalida-
tion strategies in wireless environments,” IEEE Trans. Parallel Dis-
trib. Syst., vol. 12, no. 8, pp. 789–807, Aug. 2001.
Xiaobin Tan received the BS and PhD degrees
[17] E. J. O’neil, P. E. O’neil, and G. Weikum, “The LRU-K page
from the University of Science and Technology of
replacement algorithm for database disk buffering,” ACM SIG-
China (USTC), Hefei, China, in 1996 and 2003,
MOD Rec., vol. 22, no. 2, pp. 297–306, 1993.
respectively. He is currently an associate professor
[18] D. Lee et al., “LRFU: A spectrum of policies that subsumes the
with the School of Information Science and Tech-
least recently used and least frequently used policies,” IEEE Trans.
nology, USTC. His research interests include net-
Comput., vol. 50, no. 12, pp. 1352–1361, Dec. 2001.
work performance optimization and information
[19] W. F. King, “Analysis of paging algorithms,” in Proc. IFIP Congr.,
security.
1971, pp. 485–490.
[20] P. Flajolet, D. Gardy, and L. Thimonier, “Birthday paradox, cou-
pon collectors, caching algorithms and self-organizing search,”
Discrete Appl. Math., vol. 39, no. 3, pp. 207–229, 1992.
[21] A. Dan and D. Towsley, “An approximate analysis of the LRU and
FIFO buffer replacement schemes,” ACM SIGMETRICS Perform. Jian Yang (Senior Member, IEEE) received the
Eval. Rev., vol. 18, no. 1, pp. 143–152, Apr. 1990. BS and PhD degrees from the University of Sci-
[22] E. J. Rosensweig, J. Kurose, and D. Towsley, “Approximate models ence and Technology of China (USTC), Hefei,
for general cache networks,” in Proc. IEEE INFOCOM, 2010, pp. 1–9. China, in 2001 and 2006, respectively. From 2006
[23] H. Che, Y. Tung, and Z. Wang, “Hierarchical Web caching sys- to 2008, he was a postdoctoral scholar with the
tems: Modeling, design and experimental results,” IEEE J. Sel. Department of Electronic Engineering and Infor-
Areas Commun., vol. 20, no. 7, pp. 1305–1314, Sep. 2002. mation Science, USTC. Since 2008, he has been
[24] C. Fricker, P. Robert, and J. Roberts, “A versatile and accurate an associate professor with the Department of
approximation for LRU cache performance,” in Proc. 24th Int. Tele- Automation, USTC. He is currently a professor
traffic Congr., 2012, pp. 1–8. with the School of Information Science and Tech-
[25] V. Martina, M. Garetto, and E. Leonardi, “A unified approach to nology, USTC. His research interests include
the performance analysis of caching systems,” in Proc. IEEE Conf. future network, distributed system design, modeling and optimization,
Comput. Commun., 2014, pp. 2040–2048. and multimedia over wired and wireless and stochastic optimization. He
[26] X. Tang, H. Chi, and S. T. Chanson, “Optimal replica placement was the recipient of Lu Jia-Xi Young Talent Award from the Chinese
under TTL-based consistency,” IEEE Trans. Parallel Distrib. Syst., Academy of Sciences in 2009.
vol. 18, no. 3, pp. 351–363, Mar. 2007.
[27] M. Amadeo, C. Campolo, G. Ruggeri, G. Lia, and A. Molinaro,
“Caching transient contents in vehicular named data networking: Xiaofeng Jiang (Member, IEEE) received the BS
A performance analysis,” Sensors, vol. 20, no. 7, 2020, Art. no. 1985. and PhD degrees from the University of Science
[28] S. Vural, N. Wang, P. Navaratnam, and R. Tafazolli, “Caching and Technology of China (USTC), Hefei, China,
transient data in internet content routers,” IEEE/ACM Trans. in 2008 and 2013, respectively. He is currently an
Netw., vol. 25, no. 2, pp. 1048–1061, Apr. 2017. associate research fellow with the School of Infor-
[29] E. Leonardi and G. L. Torrisi, “Least recently used caches under mation Science and Technology, USTC. His
the shot noise model,” in Proc. IEEE Conf. Comput. Commun., 2015, research interests include partially observable
pp. 2281–2289. arkov decision processes (POMDPs), big data
[30] S. M. Ross et al., Stochastic Processes. New York, NY, USA: Wiley, and higher-order tensor analysis, future network
1996. performance analysis and modeling, and cogni-
tive wireless and radar counter measures.
Quan Zheng received the BS degree in produc-
tion process automation from the Dalian Univer-
sity of Technology, Dalian, China, in 1992, and " For more information on this or any other computing topic,
the MS degree in automatic control theory and please visit our Digital Library at www.computer.org/csdl.
application and the PhD degree in computer
software and theory from the University of Sci-
ence and Technology of China, Hefei, China, in
1995 and 2003, respectively. He is currently an
associate professor with the Department of
Automation and the deputy director of Labora-
tory for Future Networks. His research interests
include video semantic retrieval, media content distribution, video qual-
ity detection, and future networks.
Authorized licensed use limited to: UNIVERSITAS TELKOM. Downloaded on December 05,2023 at 17:39:42 UTC from IEEE Xplore. Restrictions apply.