External Routing Mechanisms for Hierarchical Mobile Networks
3. The replacement policy
6.2.3 Service adaptation strategies
}
victim_sender= select_victim_sender();
while (there are marked mails whose sender is victim_sender and free space is not enough ){
victim_mail= select_victim_mail (victim_sender, "marked");
delete(victim_mail);
add_free_space(victim_mail);
} }
Figure 6.16: Algorithm of three-level replacement policy by taking e-mail for example.
field we mentioned in the beginning of Section 3. For example, if the most important entry of e-mail is the sender, sender will be the second criterion after user decision. The mobile device needs to record the access history of each sender in the CAHB. The access history includes the last access timestamp and the access count of content information sent from the certain sender. These two values can obtain a replacement weight value for each sender by
R WeightValue(Sender
Figure 6.16 is an algorithm of replacement in e-mail services if we apply two criteria, user decision and sender.
6.2.3 Service adaptation strategies
There are lots of personal information services, and they may have different characteristics. In this section, we give an example of how to divide a row of information into primary, content, and ignored parts for three kinds of personal information services, which are e-mail, calendar, and business cards. Former studies [52, 30, 18] show that how people access their personal information is highly correlated with the owner and the time of information. We can briefly conclude our observations below and follow these observations to list examples of the abstrac-tion results in Table 6.2. Since the storage space of PDAs is larger than that of cell phones,
Table 6.2: Example of abstraction for e-mail, calendar, and business card.
E-mail Calendar Business card
Cellular
phone PDA Cellular
phone PDA Cellular
phone PDA
Primary Sender, Sender, Participant, Participant Owner, Owner, information Title Title, Date/time Date/time Phone Phone,
Date/time subject, Email addr.
place
Content Date/Time, Mail body, Subject, Description Email addr, Affiliation, information Mail body, Attachment Place, Affiliation, H/O addr.,
Title of
at-tachment Description H/O addr., URL,
URL, Instant msg
Instant msg
Ignored Version, Version, Reminder, Reminder, Birthday, Birthday,
information Path, Path, ... ... Nickname, Nickname,
... ... ... ...
we can put more information in primary parts to decrease the probability for users to issue dominant requests for content information.
Observation a. Important contacts usually have higher longevity and reciprocity, and are more recent than unimportant contacts.
Observation b. Important contacts usually are in the same affiliation.
Observation c. Users’ decisions to delete e-mails usually depend on who is the sender, such as the number of previous mails sent by the sender and the interaction frequency with the sender.
6.2.4 Performance evaluation
We evaluate the performance based on our prototype in this section. We first study the number of mails an email system can store in a mobile device with or without applying PIH.
In this section, we demonstrate the effectiveness of the HPIM scheme on the number of mails available to a user. As mentioned earlier, the memory space of an HPIM e-mail system is divided into four blocks: IIMB, PIB, CB, and CAHB. An HPIM device keeps only the primary information of mails, instead of storing the full header of each mail, as an ordinary
Table 6.3: Entry sizes of IIMB, PIB and CAHB of a cellular phone.
memory block entry size
IIMB 8 bytes
PIB 40 bytes
CAHB 21 bytes
CB Flexible
PIM e-mail system does. Therefore it has more space for PIB, IIMB, and CAHB to fulfill the virtual memory concept of HPIM. In an HPIM mail system, the number of CAHB entries is the same as the number of mails having their contents stored in CB. However PIB and IIMB can have as many entries if there is still available spare memory available. The total entry number of PIB and IIMB is the number of mails accessible by an HPIM user. In order to focus on the effects of PIB and IIMB memory allocation on the perceivable mail numbers, we assume the number of mail contents stored in an HPIM mobile device is the same as that in an ordinary PIM device in the following analysis. Furthermore, the average size of a mail is set as 40K bytes with a full header of 1K bytes. The entry sizes of PIB, IIMB, and CAHB of an HPIM mobile phone are 40, 8, and 21 bytes, respectively, as shown in Table 6.3.
We first simply set the size of IIMB as zero and study the effects of PIB by varying the memory sizes reserved for the e-mail system. When we set IIMB to zero, and we can calculate the number of mails accessible by an HPIM e-mail system as follows.
N
v
=
Memory Space (Avg. size of contentNCB
+Entry size of CAHBNCB )
Entry size of PIB
where Nv is the virtual number of mails accessible and NCB the number of mail contents stored in an HPIM e-mail system.
Figure 6.17 shows the number of mails stored in an ordinary PIM e-mail system and the number of mails accessible in an HPIM e-mail system with various sizes of memory space.
In Figure 6.17 (a), we set the size of IIMB as zero, and we can observe that the number of mails accessible in an HPIM e-mail system is much larger than the number of mails stored in an ordinary PIM system. In figure 6.17 (b), we studyNv under different percentage of IIMB and PIB. We can observe thatNv increases as the percentage of IIMB increases since the entry size of IIMB is much smaller than that of PIB. However, this may also influence the hit ratios.
Figure 6.18 gives us an idea of how the percentage of IIMB influences miss ratios of both
0 1000 2000 3000 4000 5000 6000
1 2 3 4 5 6 7 8 9 10
Memory space (MB)
with PIH without PIH
Number of mails in mobile device
0 500 1000 1500 2000 2500 3000
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
IIMB/(IIMB+PIB)
All PIB IIMB
Number of mails in mobile device
(a)
(b)
Figure 6.17: Effect of information hierarchy on number of mails with different memory space sizes.
0 20 40 60 80 100
1 2 3 4 5 6 7 8 9 10
Memory space (MB)
0.2 0.5 0.8
IIMB miss rate (%)
0 20 40 60 80 100
1 2 3 4 5 6 7 8 9 10
Memory space (MB)
0.2 0.5 0.8
PIB miss rate (%)
(a)
(b)
Figure 6.18: Effect of the size of IIMB on miss rates of PIB and IIMB.
0 20 40 60 80 100
0.2 0.5 0.8
IIMB/(IIMB+PIB)
PIB IIMB
miss rate (%)
Figure 6.19: Miss rates of PIB and IIMB with different percentage of the size of memory blocks and fixed number of mails.
IIMB and PIB. In general, when the memory space increases, miss rate of both IIMB and PIB decreases. If we increase the portion of IIMB, it will cause a decrement of IIMB miss rate but an increment of PIB miss rate. However, we can find in Figure 6.18 (b) that if the memory space is big enough (7 MB in Figure for example), the miss rate of IIMB is zero. Therefore, we can arrange larger size of memory space for PIB when memory space is large enough to decrease the miss rate of PIB and thus decrease information access time.
We can also observe from Figure 6.19 that if we spare some memory space from PIB to IIMB, it will slightly increase the miss rate of PIB but can improve lots of hit ratio in IIMB.
The reason is because the entry size of IIMB is much smaller than that of PIB. Therefore, a small sacrifice of size of PIB will greatly improve the hit ratio of IIMB and increase the number of virtual mails in mobile devices.
Besides the size of IIMB and PIB may influence miss rate and access time, different infor-mation access behaviors may also influence the performance of PIM system. We generate four access behaviors to study how they effect on the performance. The First one is random ac-cess, which access information randomly. The second one is Sender locality acac-cess, in which users have higher probability to access some information sent by particular group of senders.
The third one is Time locality access, in which users have higher probability to access newer information. The last one is Sender and time locality, which combines characteristics of both time locality access and sender locality access behaviors.
Figure 6.20 illustrates the results how different access behaviors influence the miss rate of IIMB, PIB, and CB. Figure 6.20 (a) and (b) are miss rates of PIB and IIMB respectively. Since we apply sequential replacement strategy for both PIB and IIMB, time is the most critical factor to decide which information should be replaced. As a result, access behaviors with time locality have lower miss rate in PIB and IIMB. On the contrary, user-assisted replacement mechanism is applied for CB and it considers all three factors, user decision, time, and sender.
Therefore, access behaviors with sender locality have lower miss rates in CB and the fourth access behavior, Sender and time locality access, has lowest miss rate while memory space increases.
Number of mails in the information server may also influence the hit ratio of PIB and
0 20 40 60 80 100
1 2 3 4 5 6 7 8 9 10
Memory space (MB)
Random Sender Loc.
Time Loc. Sender & Time Loc.
PIB miss rate (%)
0 20 40 60 80
1 2 3 4 5 6 7 8 9 10
Memory space (MB)
Random Sender Loc.
Time Loc. Sender & Time Loc.
IIMB miss rate (%)
30 40 50 60
1 2 3 4 5 6 7 8 9 10
Memory space (MB)
Random Sender Loc.
Time Loc. Sender & Time Loc.
CIB miss rate (%)
(a)
(b)
(c)
Figure 6.20: Effect of different access behaviors on miss rates of PIB , IIMB, and CIB.
0 20 40 60 80 100
500 1000 2000 3000 4000 5000 6000
Number of mails
0.2 0.5 0.8
PIB miss rate (%)
0 20 40 60 80
500 1000 2000 3000 4000 5000 6000
Number of mails
0.2 0.5 0.8
IIMB miss rate (%)
(a)
(b)
Figure 6.21: Miss rates of PIB and IIMB with different percentage of the size of memory blocks.
0
Time Loc. Sender & Time Loc.
Access time (sec.)
(a)
(b)
Figure 6.22: Effect on average access time.
IIMB. When there are more mails in the information server, the probability of replacement increases since the memory space of mobile device is fixed and the number of physical mails can be stored is also limited. Figure 6.21 shows the effect of number of mails in the server.
The most important performance result we need to study is average access time of infor-mation. Figure 6.22 demonstrates the effects of IIMB size and access behaviors on average access time. In Figure 6.22(a), we can observe that users needs to wait for at least ten seconds for accessing a mail, whereas users only need to wait for less than one second if applying proposed PIH architecture and polices. In addition, the smaller the size IIMB occupies, the shorter average access time is. The reason is that miss penalty of PIB is greater than that of IIMB. The decrement of size of IIMB will increase the hit ratio of PIB and thus decrease the average access time. In Figure 6.22(b), we can see that access behaviors with sender and time locality has the shortest access time. Since the results of prior research that personal informa-tion access behaviors follows both time and sender locality, the proposed accessing policies can improve the performance of personal information management.
6.3 Summary
In this chapter, we have proposed a load-based scheduling scheme in a two-tier network archi-tecture. The presented scheduling scheme makes use of three proxies to utilize the bandwidth of two-tier networks more effectively. Simulation results show that the presented scheme can help NEMOs in providing services with high completion rate and/or short average waiting time.
We also proposed the architecture and the accessing policies of PIH, and implemented a prototype of e-mail hierarchical system in this paper. The hierarchical architecture of PIH enables users to access and manage their personal information through mobile devices at any-time in any place. The proposed specialized accessing policies consider all characteristics of different users, services, and devices. Different users can specify their personal preferences in the memory card. Different services have the proper abstraction strategies to reduce the size of packets to be transferred to or stored in the mobile devices. Different devices run different algorithms according to their storage space and processing power. The proposed architecture
and policies of PIH can help users flexibly manage their personal information on multiple de-vices and utilize the memory of all dede-vices efficiently. In the future, we wish to implement a PIH system on a real mobile phone and study mechanisms to reduce the miss rate.