A Chapter Preloading Mechanism for E-Reader in Mobile Environment
Chen-Yuan Chuanga,b,∗, Yi-Bing Linb, Zhihao Julie Renc
aTelecommunication Laboratories, Chunghwa Telecom Co., Ltd., Taipei, Taiwan
bDepartment of Computer Science, National Chiao-Tung University, Hsinchu, Taiwan
cTelecom SudParis, Evry, France
Abstract
E-book is a fast growing service for mobile Internet. A user reads e-books through an e-reader that could be a dedicated reading device or a mobile application on the smartphones or the tablets. With content protection of Digital Right Management (DRM), e-books take longer chapter loading time due to necessary data decryption. This paper propose a mechanism that preloads a small number of chapters into the buffer for DRM-protected e- books. When the user reads a chapter in the buffer, our solution reduces the user waiting time without affecting existing content protection. We also analyze the buffer replacement cost and effects of the limited buffer resource.
Keywords:
buffer management, e-book, e-reader, preload, waiting time, Zipf’s law
1. Introduction
Nowadays, e-book service has been deployed in many countries. A user reads e-books through an e-reader that could be a dedicated reading device or a mobile application on the smartphones or the tablets [2, 3]. Most studies and surveys focused on e-reader features and functionalities [16, 23]. In particular, ePub [11] has become a dominant standard of e-book format.
We observe that while the e-reader performs data loading, pagination, rendering, and page turning, the waiting time directly influences the user
∗corresponding author
Email addresses: [email protected] (Chen-Yuan Chuang), [email protected] (Yi-Bing Lin), [email protected] (Zhihao Julie Ren)
*Manuscript (including abstract) Click here to view linked References
Figure 1: Files in a Sample ePub.
experience. As [20] demonstrated, the e-book chapter loading is a major procedure that consumes CPU time and memory resource, especially if the Digital Right Management (DRM) mechanism is involved. We have con- ducted general e-book user experience survey in the library of National Chiao- Tung University; the details are given in [19]. However, user experience of preloading to speed up encryption/decryption has not been investigated in the literature.
For open content, the e-reader can simply load the whole e-book from storage or file system [7]. On the contrary, DRM-protected e-books must be encrypted in order to avoid illegal copying [6, 10, 13]. To access DRM- protected e-book content, the e-reader requests a key and then spends addi- tional CPU time and memory resource on data decryption when loading the required chapter into the buffer. Consequently, everytime the user turns to the next or jumps to another chapter, he/she must wait for a long response time until the new chapter is loaded.
Figure 1 depicts the extracted files from a sample ePub. Rather than text-level compression [14], an ePub file basically is a zip-archived file, and the mimetype file (Line 1) indicates that the zip file follows the ePub-specific file structure. The ePub works with XML and Web technologies. XML files
describe the metadata of the e-book, including book title, author, Table of Contents, and manifest file (Lines 3 and 7).
In most of e-books, each chapter is encapsulated in an XHTML file (Lines 8 and 10). Like a Web page, the XHTML file might also reference some pictures and CSS files (Lines 9 and 12). File signatures and encryption (Lines 4-5) are optional and proprietary, but note that the DRM mechanism, e.g., ADEPT (Adobe Digital Experience Protection Technology) [1], only encrypts the chapter files rather than the whole ePub file.
Practically, the e-readers are built upon the web-browser components with the common APIs: loadURL() and loadData() [12, 17, 22]. The former loads the chapter contents from web sites or files, and the latter loads a single string as the chapter contents. Therefore, to feed the plaintext in secure paths (e.g., memory/buffer rather than files), the e-reader may host a local ServerSocket which decrypts the chapter files and mimics as a web site. Alternatively, the e-reader decrypts the chapter files and composes all materials into a string [15]. As a result, the DRM-protection induces extra user waiting time on chapter loading as compared with DRM-free scenario. In Chunghwa Telecom [8], we have conducted experiments which indicate that with a cache (buffer) mechanism, an e-reader with DRM-protection performs as fast as DRM-free one does.
This paper proposes a buffer management scheme that preloads a small number of chapters into the limited buffer. If the user intents to read a chapter in the buffer, this proposed mechanism reduces chapter loading time and thus improves the user experience.
2. Chapter Preloading Mechanism
The proposed chapter preloading mechanism uses a multi-element ring buffer [21], where each element holds one chapter of the e-book. Moreover, the e-reader loads and presents the current chapter in main thread and then preloads remaining chapters as background process. As long as the memory usage of the process is allowed on the mobile device, the e-reader can de- crypt the ciphertext from files and preload the plaintext into buffer chapter by chapter. The concept of our preloading mechanism is similar to cache management described in [5, 9], but the execution is very different.
Our mechanism and system model can apply to most e-readers that pro- vide Next and Previous buttons, and support chapter jumping from the Table Of Contents (TOC) of the e-book. Although some of them can further
2 First Author et al. ETRI Journal, Volume x, Number y, Date
For the description purpose, we define two operators ! and ! as follows:
n ! j = (n + j) mod N and n ! j = (n – j) mod N where j is the number of chapters the user jumps forward or backward from chapter n.
To accommodate the ! and the ! operators on the chapter sequence [1,N]={1, 2, …, N}, we define
a,b = [ ]a,b if a! b
[ ]a, N " 1,b[ ] if a> b
#$
%
&% (1)
When the reader jumps from chapter n to another chapter n*, we call it
• a forward jump if n*! n, n " N / 2#$ %& .
• a backward jump if n*! n! N / 2("# $% &1), n!1 .
The MLRO ranking of chapter n* with respect to n is mathematically defined as follows:
and j will be
j= kn(n*) 2
!
"!
#
$# (2)
After jumping to chapter n*, the MLRO ranking of all chapters is reset with respect to the new current chapter. We always load the chapters with the smallest MLRO ranks into the buffer. Note that in order to reduce the user waiting time, preloading the current chapter is the highest priority task and the remaining chapters are preloaded by background task with lower priority.
Let B be the buffer size, then chapter i is in the buffer (i.e. 1 ӊkn(i) ӊ B) when
i!
n! B" 1 2
#$% &
'(,n⊕ B" 1 2
#$% &
'( if B is odd n! B
2" 1
#$% &
'(,n⊕ B 2
#$% &
'( if B is even )
* ++ , ++
(3)
2. Chapter Replacement Policy
We consider that the limited buffer size cannot preload the whole e-book (N >B), so that some chapters in the buffer may be replaced. The number of chapters to be replaced is derived in the following Theorem and illustrated in Figure 2.
Theorem 1. Let NL(j) be the number of replaced chapters in the buffer. Then
NL( j)=
j if j< B and j < N ! B N! B if N ! B " j < B B if B" j " N ! B
#
$%
&%
(4-1) (4-2) (4-3)
1 2 3
4 5 6 N
n*
7 1 2
3 4 5 6
N n*
7
1 2
3 4 5 6 N
n
7
n* = 3!
j = 2!
r = 7!
l = 0!
1 2 3
4 5 6
N n
n1 n*
7
1 2
3 4 5 6 N
n
7
1 2 3
4 5 6 N
n*
7 B = 5!
N = 14!
N-B = 9!
n = 1!
n* = 7!
j = 6!
r = 0!
l = 0!
n* = 7!
j = 6!
r = 3!
l = 1!
B = 9!
N = 14!
N-B = 5!
n = 1!
n* = 3!
j = 2!
r = 3!
l = 0!
(a). Initial state with smaller B
(b). One-side overlapping from (a)
(c). No overlapping from (a)
(d). Initial state with larger B
(e). One-side overlapping from (d)
(f). Both-side overlapping from (d)
Fig. 2. Buffer Replacement Policy. (The internal ring with N elements represents the e-book. A portion the external ring with B elements represents the buffer.)
Chapter number i 1 … n - j … n - 1 n n + 1 … n + j … N
MLRO rank kn(i) … 2j +1 … 3 1 2 … 2j …
Fig. 1. Most Likely Reading Order (MLRO).
kn(n*)= kn(n! j) = 2 j for a forward jump kn(n!j)= 2 j +1 for a backward jump
"
#$
Figure 2: Most Likely Reading Order (MLRO).
jump back to a recently read chapter, we will not cover the Return/Back and Bookmark in this paper. It is expected that a dedicated element to cache the last-read chapter may improve performance and chapters with bookmarks have higher probability to be jumped to as well, and our model can easily be extended to accommodate such a feature.
2.1. Notation and Definitions
Let N be the number of chapters in a given e-book. We first rank all chapters according to its Most Likely Reading Order (MLRO) as shown in Figure 2. MLRO is established based on common reading habits, i.e., the reading order is sequential and chapters close to the current reading one have higher probabilities to be read in the near future.
Let n be the current chapter (1 ≤ n ≤ N ), and kn(i) be the MLRO rank of any Chapter i with respect to n. In Figure 2, the gray-colored box indicates Chapter n, which has the rank kn(n) = 1. Rank 2 is assigned to the next chapter in the sequential order: kn(n + 1) = 2, rank 3 is assigned to the previous one, i.e., kn(n − 1) = 3, and so on. We note that user’s “backward”
reading may not exactly follow the ranking. However, this ranking does capture that observed in our survey [19].
For the description purpose, we define two operators ⊕ and as follows:
n ⊕ j = (n + j) mod N and n j = (n − j) mod N
where j is the jump distance (in number of chapters) forward or backward from Chapter n. In other words, Chapter 1 jumps forward 1 chapter to Chapter 2, and Chapter 2 jumps backward 2 chapters to Chapter N.
Because we transform the chapter sequence [1, N ] = 1, 2, . . . , N into a ring structure, we extend the definition of chapter sequence as follows:
ha, bi = [a, b] if a ≤ b
[a, N ] k [1, b] if a > b (1) For example, h2, 5i means Chapter 2 to Chapter 5, and h5, 2i means Chapter 5 to Chapter N and Chapter 1 to Chapter 2.
Basically, the e-reader can jump from Chapter n to another Chapter n∗ within N/2 chapters, and we call it
• a forward jump if n∗ ∈
n, n ⊕ N 2
, or
• a backward jump if n∗ ∈
n N 2
− 1
, n 1
.
In Figure 2, assume that N = 11 and n = 6. If n∗ = 9, then it is a forward jump with j = 3, and k6(9) = 6. If n∗ = 1, then it is a backward jump with j = 5, and k6(1) = 11. It is clear that the MLRO ranking of Chapter n∗ with respect to Chapter n can be expressed as
kn(n∗) = kn(n ⊕ j) = 2j for a forward jump kn(n j) = 2j + 1 for a backward jump and the jump distance j will be
j = kn(n∗) 2
(2) Let B be the buffer size, then Chapter i is in the buffer (i.e., 1 ≤ kn(i) ≤ B) when
i ∈
n B − 1 2
, n ⊕ B − 1 2
if B is odd
n B 2 − 1
, n ⊕ B 2
if B is even
(3)
Equation (3) determines whether a chapter is in the buffer depending on B.
For example, in Figure 3(a), n = 1 and B = 5, then Chapters 13, 14, 1, 2 and 3 are in the buffer; in Figure 4(a), n = 1 and B = 9, then Chapters 11 to 5 are in the buffer.
2.2. Chapter Replacement Cost
We assume that the limited buffer size cannot preload the whole e-book (N > B), so that the e-reader must replace some victim chapters in the buffer. In Figure 3, when the user jumps to Chapter 3 (see Fig. 3(b)), a.k.a.
cache hit, there are 2 chapters in the buffer will be replaced; when the user
Figure 3: Buffer Replacement Mechanism (small B).
Figure 4: Buffer Replacement Mechanism (large B).
jumps to Chapter 7 from Chapter 1 (see Fig. 3(c)), a.k.a. cache miss, all chapters in the buffer will be replaced in this case.
After a chapter jump, the MLRO ranking is reset with respect to the new current chapter. If the MLRO ranking of a buffered chapter is larger than B, it implies that the chapter should be replaced; otherwise, the chapter stays in the buffer.
The number of chapters to be replaced is derived in the following Theo- rem, and scenarios of forward jumps are illustrated in Figure 3 and 4.
Theorem 1. Let NL(j) be the number of replaced chapters. Then
NL(j) =
j if j < B and j < N − B N − B if N − B ≤ j < B
B if B ≤ j ≤ N − B
(4a) (4b) (4c) Proof. Suppose that B is odd (we omit the similar proof for the case when B is even). From (3), the chapter sequence S in the buffer (see Fig. 3(a) and 4(a)) is
S =
n B − 1 2
, n ⊕ B − 1 2
(5) Without loss of generality, we assume that the user jumps j chapters forward (we omit the similar proof for a backward jump). After the jump, the chapter sequence S∗ in the buffer is
S∗ =
n∗ B − 1 2
, n∗⊕ B − 1 2
=
(n + j) B − 1 2
, (n + j) ⊕ B − 1 2
(6)
Let r and l be the numbers of overlapping chapters on the right-hand side and the left-hand side in S and S∗, respectively. For example, in Figure 4(c), we have r = 3 (Chapters 3, 4 and 5) and l = 1 (Chapter 11). The number of overlapping chapters is r + l. Therefore, the number of replaced chapters in the buffer is
NL(j) = B − (r + l) (7)
Now, we prove the hypothesis in two scenarios: j < B and j ≥ B.
Scenario 1: j < B (see Fig. 3(b), 4(b) and 4(c)). We first consider the overlapping chapters on the right-hand side. Since j < B, from (5) and (6), we know that the overlapping chapters on the right-hand side must be in
(n + j) B − 1 2
, n ⊕ B − 1 2
(8) From (8),
r =
n + B − 1 2
−
(n + j) − B − 1 2
+ 1 = B − j (9) Now, we consider the left-hand side: If l 6= 0, (see Fig. 4(c)), it implies that the chapters not in S must be loaded in S∗, and
NL(j) = N − B (10)
Similarly to (8), we have the overlapping chapters on the left-hand side from (5) and (6) if l 6= 0. Then we find that
(n + j) + B − 1 2
≥ n − B − 1 2
+ N ⇔ j ≥ N − B (11) Therefore, l 6= 0 if and only if j ≥ N − B. When j < N − B (see Fig.
3(b) and 4(b)), l = 0 and from (7) and (9), NL(j) is NL(j) = B − (r + l)
= B − ((B − j) + 0)
= j
(12) From (10) and (11), we prove equation (4b), and equation (4a) is also proved from (12).
Scenario 2: j ≥ B (see Fig. 3(c)). This scenario implies r = 0. We show l = 0 by contradiction. Suppose l 6= 0, then (11) must be satisfied. When j > N − B, by definition of jump, j ≤ N/2, we have
N
2 > N − B ⇔ B > N
2 (13)
Since j ≥ B, equation (13) implies that j > N/2 which contradicts with the definition of jump, i.e., j ≤ N/2. When j = N − B, B and j will be N/2 and finally l must be 0. Therefore, Scenario 2 holds if and only if B ≤ j ≤ N − B,
NL(j) = B (14)
From (14), equation (4c) is proved.
3. Analysis and Discussion
This section shows how to select appropriate B values according to the reload probability p and the replacement cost CL. Let p be the probability that the next Chapter n∗ is not in the buffer (cache miss). Let CL be the expected number of replaced chapters in the buffer due to a chapter jump.
Note that for DRM-protected e-books, the buffer of e-reader is a limited resource, and for the practical exercise, it suffices to consider B < N/2.
We derive p and CL assuming that the probability distribution of e-book chapter reading follows Zipf’s law [4, 18]. This law predicts the frequency of the chapter of MLRO rank k to appear. The Zipf’s probability of the chapter of rank k is
PZ(k) =
"
ks
N
X
i=1
1 is
!#−1
(15) where s describes the reading behavior, and more specifically, the locality of the read chapters. A large s means that the user is inclined to read neighboring chapters; on the contrary, a small s means that there is a higher probability to jump further away from the current chapter. Zipf’s law gives the probability of any chapter to be read depending on its MLRO rank.
From (15), the reload probability p is expressed as
p = X
kn(n∗)>B
PZ[kn(n∗)]
= 1 −P
1≤k≤B
"
ks
N
X
i=1
1 is
!#−1 (16)
Now we derive CL. For the practical exercise, we assume B < N/2, so Theorem 1 can be simplified as
NL(j) = j if j < B B if B ≤ j
Let C1 be the replacement cost corresponding to forward jumps. Since kn(n∗) = 2j and j ≤ bN/2c, we have
C1 = X
1≤j≤bN2c
NL(j)PZ(2j) = X
1≤j<B
jPZ(2j) + B
X
B≤j≤bN2c
PZ(2j)
(17)
Figure 5: Effect of s and B/N on p (N =20).
Let C2 be the replacement cost corresponding to backward jumps. Since kn(n∗) = 2j + 1 and j ≤ dN/2e − 1, we have
C2 = X
1≤j≤dN2e−1
NL(j)PZ(2j + 1)
= X
1≤j<B
jPZ(2j + 1) + B
X
B≤j≤dN2e−1
PZ(2j + 1)
(18)
From (17), (18), (15) and (2), the replacement cost CL is CL = C1+ C2
= X
1≤l<2B
l 2
PZ(l) + B X
2B≤l≤N
PZ(l)
!
= X
1≤l<2B
l 2
"
ls
N
X
i=1
1 is
!#−1 + B
X
2B≤l≤N
"
ls
n
X
i=1
1 is
!#−1
(19) From (16), we observe that a larger B implies a lower p. On the contrary, from (19), a larger B implies a higher CL. To select an appropriate B to
Figure 6: Effect of N and B/N on p (s=2).
Figure 7: Effect of s and B/N on CL (N =20).
balance against p and CL, we define a net cost C as follows:
C = p + αCL
where α is a factor that weights the importance of CL against p.
Figure 5 shows the reload probability p as a function of B/N and s, where N =20. The figure indicates the intuition that p decreases as B/N increases. We observe a non-trivial result that there is a knee point (the bullets in Figure 5) such that before this point, p significantly decreases as B/N increases. This phenomenon implies that after this knee point, selecting a larger buffer size does not improve p performance. Such knee effect is more significant for a larger s. Figure 5 also indicates that to achieve the same p performance, B/N deceases as s increases.
Figure 6 illustrates that to achieve the same p performance; a larger B/N is required for a smaller N . That is, a larger portion of e-book needs to be stored in the buffer for an e-book of small volume. Similar to Figure 5, a knee point can be determined on every p curve.
Figure 7 shows the loading cost for N =20 under various s values. We observe that CLincreases as the locality parameter s or the e-book size B/N increases; this effect becomes less significant with large s values.
The net cost for s=2 is illustrated in Figure 8. For α=0.6, we observe a minimum value on each C curve for N =50 and N =100, respectively. In both cases, this minimum occurs when B=3 (B/N =0.06 for N =50 and B/N =0.03 for N =100). By increasing the factor α, the protection of the e-book is higher and the minimum occurs for a smaller B value (for α=0.8, B=2). We note that the B/N value of the minimum is always smaller than that of the knee point for p.
The above performance study indicates that by slightly increasing the buffer size, the user experience is significantly improved. For example in Figure 8, by buffering two more chapters (i.e., B/N increases from 0.02 to 0.06 for N =50 and α=0.6), we can reduce the reload probability p from 0.38 to 0.17 (Figure 6), which is a 56% improvement. Our chapter preloading mechanism provides a method for e-reader designers to select the appropriate buffer size according to the requirements in term of user experience and e- book protection.
Figure 8: Effect of N , α and B/N on C (s=2).
4. Conclusions
This paper proposed a chapter preloading mechanism for e-reader that significantly enhances the user experience by slightly increasing the buffer resource with the lower reload probability. Furthermore, we also guided the e-reader developers to allocate the appropriate buffer resource to balance the cache-hit ratio and replacement cost.
Our scheme preloads more chapters in the buffer than the existing sce- nario (i.e., no preload). Since in both approaches, the decrypted chapters are loaded into the secure paths that are not accessible by other processes, our scheme is compatible with the existing DRM-protection and assures the required protection level for the e-book.
The forthcoming ePub 3.0 will enrich the e-book contents, and different reading sequences are expected for various contents. A potential future work is to extend the MLRO ranking to be more precise for various classifications of e-book and the features of e-readers. Same technique proposed in this paper can be used to model the access performance of the new reading features.
References
[1] Adobe - Reader Developer Center — Adobe Developer Connection, Available at: http://www.adobe.com/devnet/reader.html, March 25, 2012.
[2] Amazon.com: Kindle eBooks, Newspapers, Magazines, Blogs, Available at: http://www.amazon.com/, March 25, 2012.
[3] App Store - Stanza, Available at: http://itunes.apple.com/us/app/
stanza/id284956128?mt=8, March 25, 2012.
[4] L. Breslau, C. Pei, F. Li, G. Phillips, S. Shenker, Web caching and Zipf- like distributions: evidence and implications, in: Proceedings of IEEE INFOCOM’99, March 1999, pp 126–134.
[5] F.-C. Chang, H.-C. Huang, A refactoring method for cache-efficient swarm intelligence algorithms, Information Sciences 192 (2012) 39 – 49.
[6] C.-Y. Chuang, Y.-C. Wang, Y.-B. Lin, Digital right management and software protection on Android phones, in: Proceedings of the 71st Ve- hicular Technology Conference (VTC 2010-Spring), 2010.
[7] Freda: the free ebook reader. Available at: http://www.turnip.demon.
co.uk/jim/freda/, March 25, 2012.
[8] HamiBook (in Chinese) — Chunghwa Telecom Co., Ltd., Available at http://bookstore.emome.net/, March 25, 2012.
[9] W.-S. Han, W.-K. Loh, K.-Y. Whang, Type-level access pattern view:
Enhancing prefetching performance using the iterative and recursive pat- terns, Information Sciences 180 (2010) 4118-4135.
[10] G. He, Analysis of e-book security, Publications in Telecommunications Software and Multimedia (2001), Available at: http://www.tml.tkk.
fi/Studies/Tik-110.501/2001/papers/he.pdf, March 25, 2012.
[11] International Digital Publishing Forum — Trade and Standards Or- ganization for the Digital Publishing Industry, Available at http:
//idpf.org/, March 25, 2012.
[12] iOS Developer Library: UIWebView Class Reference, Available at: http://developer.apple.com/library/ios/#documentation/
uikit/reference/UIWebView_Class/Reference/Reference.html, March 25, 2012.
[13] Q. Liu, R. Safavi-Naini, N. Sheppaed, Digital rights management for content distribution, in: Proceedings of the Australasian information security workshop, 2003.
[14] M. A. Martinez-Prieto, J. Adiego, P. Fuente, Natural language compres- sion on edge-guided text preprocessing, Information Sciences 181 (2011) 5387-5411.
[15] L. Masinter, The ”data” URL scheme, RFC 2397, IETF, August 1998.
[16] L. M. Moore, At your leisure: assessing ebook reader functionality and interactivity, Master Thesis, University College, London, UK, 2009.
[17] MSDN Library: WebBrowser Class (Microsoft.Phone.Controls), Avail- able at: http://msdn.microsoft.com/en-us/library/ff402601(v=
vs.92).aspx, March 25, 2012.
[18] M. E. J. Newman, Power laws, pareto distributions and Zipf’s law, Con- temporary Physics 46 (2005) 323–351.
[19] Z. J. Ren, Performance evaluation and user experience of e-book reader, Master Thesis, National Chiao-Tung University, Taiwan, 2011.
[20] Z. J. Ren, Y.-B. Lin, N. Crespi, Evaluation of e-book reader on mo- bile devices, in: Proceedings of the 15th Mobile Computing Workshop, Taichung, Taiwan, 2010.
[21] P. K. C. Tse, Multimedia Information Storage and Retrieval: Techniques and Technologies, IGI Global, 2008.
[22] WebView — Android Developers, Available at http://developer.
android.com/reference/android/webkit/WebView.html, March 25, 2012.
[23] R. Wilson, E-book readers in higher education, Journal of Educational Technology Society 6 (2003) 8–17.
