Usage­Based DHCP Lease Time Optimization
Manas Khadilkar, Nick Feamster, Matt Sanders, and Russ Clark
College of Computing, Georgia Tech
ABSTRACT
The Dynamic Host Configuration Protocol (DHCP) is used to dy-
namically allocate address space to hosts on a local area network.
Despite its widespread usage, few studies exist on DHCP usage
patterns, and even less is known about the importance of setting the
lease time (the time that a client retains ownership over some IP
address) to an appropriate value. Lease time can greatly affect the
tradeoff between address space utilization and the number of both
renewal messages and client session expirations. In this paper, us-
ing a DHCP trace for 5 weekdays from the Georgia Tech campus
network, we present the largest known study of DHCP utilization.
We also explore how various strategies for setting lease times can
dramatically reduce the number of renewals and expirations with-
out prohibitively increasing address space utilization.
Categories and Subject Descriptors
C.2.3 [Computer Communication Networks]: Network opera-
tions – network management
General Terms
Algorithms, Design, Management, Measurement
Keywords
DHCP, optimization, network management, usage
1. Introduction
The Dynamic Host Configuration Protocol (DHCP) [2] allows
networks to automatically assign IP addresses to clients and set up
various configuration parameters. Many networks use DHCP to
reduce client configuration when allocating IP addresses, particu-
larly when the network comprises many mobile or intermittently
connected clients. DHCP is useful for managing IP address space
allocation for networks where the total number of users outstrips
the total number of concurrent users. For example, the Georgia
Tech “LAWN” (Local Area Walkup and Wireless Network) [3]
serves about 6,000 unique IP addresses per day, even though the
maximum number of concurrent users never exceeds 2,700; this
dynamism can preserve scarce address space if IP addresses are
allocated and reclaimed properly.
DHCP’s effect on IP address space consumption depends on how
it is configured to reclaim IP addresses. The DHCP server’s lease
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
IMC ’07 October 24–26, 2007, San Diego, CA.
Copyright 2007 ACM 978­1­59593­908­1/07/0010 ...$5.00.
time setting controls how often a DHCP server reclaims an allo-
cated IP address: if a DHCP server does not receive a renew re-
quest within the lease time, it expires the client’s lease and reclaims
its IP address for possible allocation to other clients. Ideally, a
DHCP server could reclaim an IP address immediately after a client
“leaves” the network. Unfortunately, DHCP does not mandate that
clients explicitly notify the server when they leave, so servers com-
monly issue a single lease time for all clients1; each client issues
periodic “renew” requests to keep its lease active.
As a result, network operators must set the DHCP lease time
judiciously. A lease time that is too large prevents a server from
reclaiming IP addresses being used by clients that are no longer
present on the network, unnecessarily wasting valuable IP ad-
dresses in a sometimes scarce address pool. On the other hand, set-
ting a lease time that is too small can introduce substantial DHCP
broadcast traffic from clients, which can overload switches and
send unnecessary traffic to thousands of clients. Setting short lease
times also unnecessarily expires client leases if the clients are mo-
bile and leave for only short period of time. On many networks
that use dynamic addressing, including the Georgia Tech campus
network, DHCP leases are tied to authentication. Thus, a client
whose lease expires must re-authenticate—a manual, inconvenient,
and time-consuming process.
Today, configuring DHCP lease times is a black art: network op-
erators typically choose a small fixed lease time and after carefully
observing the network over some time period, cautiously increase
the lease time if they think the network can sustain it. This paper
presents three main contributions.
First, we present a measurement study of DHCP utilization on
the Georgia Tech Local-Area Walkup/Wireless Network (LAWN),
a dynamic campus-wide network with about 6,000 unique users per
day (Section 4). We find that, for a typical day, the median client
session time is about 75 minutes, and that more than 30% of clients
return within 60 minutes after leaving the network.
Second, we present an emulation technique that can help oper-
ators evaluate the effects of longer DHCP lease times on address
space consumption and DHCP renewals and expires, given only
an existing DHCP usage trace from an operational network (Sec-
tion 5). We use this method to estimate utilization, expiry, and re-
newals on the LAWN (Section 6) and find that tripling the current
lease time (to 90 minutes) reduces re-logins by almost 23%, while
only increasing address space utilization by 14%.
Third, we explore two alternative strategies for setting lease
times that further reduce the number of renewals and expires for
a given utilization of available address space: single adjustment
and exponential (Section 7). In the single adjustment strategy, the
DHCP server sets one lease time when a client first appears and a
second lease time for all subsequent renewals from that client. The
exponential strategy doubles the lease time for each renewal, up to
a maximum possible lease time. We find that exponential backoff
1Many DHCP servers allow lease times to be configured for specific clients,
but large campus and enterprise networks typically use a single setting.

can reduce the number of client expirations by 27% while increas-
ing peak address space utilization by only 17%.
Although our measurement results reflect usage patterns that are
specific to the Georgia Tech network, both (1) the emulation tech-
nique we present for evaluating the effects of different lease time
settings and (2) the strategies for determining the best lease time
settings for a particular network are applicable to any network that
uses DHCP to dynamically allocate addresses to clients. We be-
lieve that our techniques for studying various lease times will prove
useful in any setting where operators must confront the tradeoffs
imposed by DHCP lease time settings.
The rest of the paper is organized as follows. Section 2 presents
background on DHCP and previous studies. Section 3 describes
our measurement setup. Section 4 describes usage statistics for
clients on the Georgia Tech campus network; these measurements
provide insight into client activity patterns and can prove useful
for setting DHCP lease times. Section 5 describes our algorithm
for emulating the effects of longer lease times on address space
utilization. Section 6 shows our analysis of the effects of various
lease time settings, and Section 7 shows the effects of using several
dynamic strategies for setting lease times. Section 8 concludes.
2. DHCP Background and Related Work
We briefly summarize DHCP operation and related work.
2.1 DHCP Background
The Dynamic Host Configuration Protocol (DHCP) [2] auto-
mates host configuration on a network (e.g., IP address, subnet
mask, default gateway, etc.) and helps network administrators dy-
namically allocate IP addresses to clients from a pool of IP ad-
dresses. Because it both automates client configurations and helps
operators dynamically assign addresses from a relatively small pool
of IP addresses, DHCP has gained widespread usage.
DHCP has four common message types. When a client first at-
tempts to obtain configuration, it broadcasts a discover message
to the subnet’s broadcast address (or to the broadcast destination
255.255.255.255). A DHCP server that is listening on that sub-
net then replies with an offer message, which contains the client’s
MAC address, the offered IP address, the duration of the lease, and
other specific configuration information. Upon receiving an offer,
a client then replies to the offer message with a request for one of
the offers it received from a DHCP server on the subnet; a DHCP
server responds to a client request with an acknowledgment.
A DHCP server reclaims a client’s IP address when the client’s
lease expires (we also say that a client’s “session” expires when its
lease expires). An active client (i.e., one that is present on the net-
work) periodically “renews” its lease by sending DHCP REQUEST
messages at an interval of half the lease time, at which point the
client’s lease is extended for the duration of the full lease period.
For example, if a client’s lease time is 30 minutes, it will issue
DHCP requests every 15 minutes. We describe in Section 3.2 how
we use these renewal messages to estimate when a client is active.
2.2 Related Work
Despite its widespread use, there is relatively little understanding
about how to configure DHCP. In 1995, Perkins and Luo studied
how to use DHCP in conjunction with mobile IP to support auto-
configuration of mobile computers [5]; we measure the behavior
of DHCP users on a large network comprised primarily of mobile
hosts (e.g., laptops). Previous work has studied wireless usage on
campus networks [4] but focused on other issues (e.g., mobility),
rather than DHCP.
Birk et al. studied DHCP client behavior on two networks: one
with a dynamic pool of 420 addresses and another with a pool of
about 200 addresses [1]. They found diurnal and weekly trends in
active usage patterns, which is consistent with our findings. Birk
et al. acknowledge that choosing the size of the dynamic address
pool and setting the default lease time allocate to clients can greatly
affect DHCP “performance”. This paper offers the first study of the
effects of various DHCP client lease time settings.
3. Measurement and Analysis Setup
We describe the campus network, our DHCP measurements, and
our methods for estimating client activity.
3.1 Environment and Data
We measured DHCP behavior on the Georgia Tech Local Area
Walkup/Wireless Network (LAWN), a campus-wide local-area net-
work with a peak usage of more than 2,500 concurrent users and
a usable address space of about dynamically allocatable 4,000 IP
addresses. The LAWN comprises 1,000 access points and 2,800
“wired” network ports (most of which are not active at any point in
time) across the campus. Because the LAWN is a single virtual lo-
cal area network (VLAN), all client discover and request broadcast
messages are served by a single DHCP server.
We studied DHCP behavior using DHCP server logs from the
LAWN DHCP server collected from February 12, 2007 to Febru-
ary 16, 2007. We present the results from a five-day trace because
it represents a typical week on campus; other weeks exhibit sim-
ilar behavior. Each entry in these logs represents a single mes-
sage and indicates the type of DHCP message that was sent or
received, the time of the message, and the client’s MAC address
and an anonymized IP address. We used client MAC addresses to
uniquely identify individual clients. The LAWN is an order of mag-
nitude larger than other public DHCP studies [1], in terms of both
the size of the address pool and the number of active clients.
3.2 Estimating Duration of Client Activity
To study both basic usage patterns and the effects of changing
DHCP lease times on traffic patterns and address space utilization,
we must first understand the dynamics of client activity. Because
clients regularly issue DHCP messages, we can use the DHCP mes-
sages themselves to infer client activity. Clients issue “renew” mes-
sages at time intervals that are roughly half the lease time. The
DHCP protocol expects the clients to renew their lease when half
of the issued lease time has passed. The periodicity at which the
client renews its lease is different for different lease periods; on
the LAWN, the lease time is 30 minutes. We found that over 84%
of the client sessions renewed their leases in the 15th minute (for a
lease period of 30 minutes); more than 15% of the sessions renewed
sooner than 15 minutes, and only 1% renewed after 15 minutes.
DHCP renew requests allow us to estimate when the client is
active. The time window of the estimated client departure can be
determined to within the expected renewal time: a client could have
left the network anywhere between its last seen request and the
expected renewal time after that request. We assume that clients
leave within this period according to a uniform distribution and thus
estimate that a client is active for a period of 7.5 minutes since
its last request (i.e., half of the renewal period in the case of the
LAWN).
4. Usage Statistics
This section presents statistics of client activity on the LAWN.
We first examine five days of DHCP logs to determine the net-
work’s address space consumption under the current 30-minute

0 720 1440 2160 2880 3600 4320 5040 5760 6480 7200
Time (in minutes)
600
800
1000
1200
1400
1600
1800
2000
2200
2400
2600
2800
N
um
be
r o
f a
ct
iv
e
le
as
es
Figure 1: Address space utilization for five days on the Georgia Tech
LAWN campus network, with 30-minute lease times.
lease time setting. We also examine the distributions of on time
(i.e., the time that a client is active) and off time (i.e., the time be-
tween periods of client activity). These statistics serve not only as
a useful point of comparison to previous work [1], but also help us
better understand how different lease times affect client renewals
and expires.
4.1 Address Space Utilization
Figure 1 shows activity for the weekdays February 12–16, 2007.
The vertical dotted lines indicate midnight for each day of the week.
As expected, client activity on the campus network shows diurnal
and weekly trends (we note that Friday shows lower client activity,
likely due to the fact that fewer classes are held on Friday). Address
space utilization peaks at around 2 p.m. each day (at approximately
2,700 concurrent users for each weekday except for Friday) and is
lowest at around 6 a.m. (at about 600 concurrent users). We expect
that many of the 600 concurrent DHCP clients around 6 a.m. are
persistently connected via either a wireless access point or by a
wired port.
Although the usage patterns on the LAWN are similar to those
seen in previous work [1], we observe two notable differences
(other than the magnitude of the activity). First, the usage patterns
shown in Figure 1 show a drop in activity after the peak at roughly
2 p.m., but this drop in utilization slows at around 9 p.m. on Mon-
day through Thursday (though not on Friday). Birk et al. likely did
not observe this characteristic because they observed DHCP utiliza-
tion in an office building and research laboratory, where users are
less likely to become inactive and later return to the same network.
Second, usage patterns are slightly more erratic. We attribute
this characteristic to two underlying causes: (1) the existence of
more “high frequency” activity (e.g., clients disconnecting and re-
connecting as students move between classes); and (2) consider-
ably shorter client lease times (30 minutes, as opposed to 48 hours),
which preserves such high-frequency activity.
4.2 Individual Client Dynamics
In this section, we present findings about individual client activ-
ity. We studied two characteristics of individual client sessions:
• On-Time is the duration of time for which a client remains
active.
• Off-Time is the duration of time between when a client ini-
tiates a new session and the time that the client’s previous
lease expires.
10 30 100 1000
On-Time (in minutes)
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
11000
12000
13000
Cu
m
ul
at
iv
e
In
sta
nc
es
Figure 2: Cumulative distribution of client on-times.
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400
Off-Time (in minutes)
0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
6500
7000
7500
Cu
m
ul
at
iv
e
in
sta
nc
es
Figure 3: Cumulative distribution of client off-times.
The rest of this section discusses the motivation for studying each
of these statistics, explains the difficulties in estimating individual
on-time and off-time values, and presents the distribution of these
statistics for the LAWN.
On-time Statistics. Computing the on-time for each client helps us
determine how a particular lease time might reduce the number of
lease renewals, since increasing lease times reduces the frequency
at which a client must issue requests for lease renewal. Unfortu-
nately, computing the exact on-time for a particular session is diffi-
cult because the DHCP specification does not mandate that clients
issue a RELEASE when they leave the network [2], and almost
93% of clients in our trace do not issue a RELEASE when they
leave. As a result, a client could have left the network any time
between its last seen REQUEST and the time when its lease ex-
pires. In the case of the LAWN, a client could have “left” anytime
between within 30 minutes after its last observed request.
We approximate on-time for each session by using a client’s RE-
QUEST messages as an indication that the client is still active, as
described in Section 3.2: specifically, we subtract the last seen RE-
QUEST for a client from the first one seen since an expiry, and add
7.5 minutes to this value.

Algorithm 1 Generating Logs with Increased Lease Times
Client is expired
if Previous Message from Client is a RELEASE then
Log a REQUEST at the current lease time
else
Could be a live or expired client
if Client is timed out based on original lease time then
if Current time is greater than the next replay renewal time then
Log a REQUEST at the current lease time
end if
else
Client is live
if The expected “live” time of the client is greater than the next
replay renewal time then
Log a REQUEST at the next replay renewal time
end if
end if
end if
Figure 2 shows the cumulative distribution of on-times for the
client sessions for February 12, 2007, which had more than 14,000
unique sessions. About 20 percent of sessions lasted 30 minutes
or less (i.e., they issued two renew requests or fewer), and 59% of
all sessions lasted 90 minutes or less. This distribution suggests
that, subject to address utilization constraints, a significant fraction
of DHCP renew traffic could be saved by increasing lease time to
90 minutes. In fact, we see in Section 6 that increasing the lease
time to 90 minutes for all clients saves about 70% of the renew
messages seen with a lease time of 30 minutes without prohibitively
increasing peak address space utilization.
Off-time Statistics. Setting longer lease times not only reduces
renewal traffic for clients that are active, but it also prevents tem-
porarily inactive clients from being prematurely logged out. Com-
puting the off-time for each client allows us to determine how lease
time settings affect the number of times that a client that leaves for
a short period of time and then returns might experience lease-time
expiry, and, as a result, need to obtain a new lease (and, in the case
of the LAWN, need to re-authenticate).
We approximate off-time with a method that is similar to our
on-time estimation: We compute a client’s session off-time by sub-
tracting the time that we last saw a renewal message from the time
that we see another client renewal message and subtracting 30 min-
utes from this time interval. Figure 3 shows the distribution of off-
times for the client sessions on February 12, 2007; the DHCP logs
indicated about 7,000 distinct instances of DHCP off-times. About
70% of off-time instances are less than 210 minutes; in other words,
when a client leaves, it is likely to return within 210 minutes about
70% of the time. Similarly, more than 30% of off-time instances
are less than 60 minutes (i.e., clients renew their leases within 90
minutes of their last seen REQUEST message about 30% of the
time). Thus, increasing lease times to 90 minutes could signifi-
cantly reduce the number of sessions where clients are forced to
re-authenticate.
5. Emulating Longer Lease Times
Network operators typically want to set the DHCP lease time to
the longest possible interval that does not put undue pressure on the
available address space. Unfortunately, to date, operators have not
had the ability to evaluate how larger lease times might affect peak
address utilization without actually increasing the lease time on a
running network. In this section, we describe a method for helping
operators emulate the effects of longer DHCP lease times using
existing DHCP usage logs, and the underlying assumptions of the
analysis. We have implemented this method in an emulation tool,
0 360 720 1080 1440 1800 2160 2520 2880 3240 3600 3960 4320 4680 5040 5400 5760
Time (in minutes)
800
1200
1600
2000
2400
2800
3200
3600
4000
N
um
be
r o
f a
ct
iv
e
le
as
es
240 min
150 min
90 min
30 min
Figure 4: Address space utilization for 4 days for 4 lease times.
which we used in our evaluation of various lease time optimization
strategies in later sections.
The algorithm uses the original log to generate a “replay log”,
which approximates the times when the client would have sent re-
new REQUEST messages for longer lease times. This technique
generates logs which can be used to determine the effects of longer
lease times on renewal traffic, premature expires, and address space
utilization.
We first introduce the following terms:
• current time: the timestamp associated with DHCP message
in the original logs.
• replay renewal time: the time when the client is expected to
renew its lease for a given lease time; equal to the time of
the last replay message seen from the client plus half of the
given lease time.
Algorithm 1 summarizes how a replay log is generated from the
original DHCP request logs. Of course, a client’s initial REQUEST,
and any REQUEST following a RELEASE will occur at the same
time for any lease time. Lease renewals are more subtle. If a client’s
lease appears to have expired between two requests (i.e., if the orig-
inal logs show a DISCOVER before the client’s next REQUEST
or if the time difference between the client’s two consecutive RE-
QUESTs is greater than the original lease time) the algorithm first
determines whether the current time is past the time at which the
client was due for the next replay renew. If so, the algorithm logs a
REQUEST message in the replay logs. For each renew sent by the
client in the original logs, the algorithm logs a renew REQUEST
whenever it determines that the client is likely to be active” past the
time of when a renew would be scheduled in the case of the longer
lease time.
6. Effects of Increased Lease Times
Longer lease periods improve usability (especially for intermit-
tently connected clients) and reduce DHCP request traffic, but they
also increase address space utilization. Accordingly, operators
want to set the longest possible lease time that still leaves sufficient
spare address space. We use the algorithm from Section 5 to deter-
mine the effects of longer lease times on address space utilization,
DHCP renews, and premature session expires.

10 am 12 pm 2 pm 4 pm 6 pm 8 pm 10 pm 12 am
Client #1
Client #2
Client #3
Client #4
Client #5
Client #6
Client #7
Client #8
Client #9
30 min lease time
10 am 12 pm 2 pm 4 pm 6 pm 8 pm 10 pm 12 am
Client #1
Client #2
Client #3
Client #4
Client #5
Client #6
Client #7
Client #8
Client #9
90 min lease time
Figure 5: Sessions for 9 clients for 30 minute and 90-minute lease times.
6.1 Address Space Utilization
We analyzed address space utilization for different lease times.
Operators prefer that peak address space utilization not exceed
80%, to accommodate unexpected bursts in utilization and the grad-
ual growth of DHCP clients.2
Figure 4 shows the concurrent leases that would exist on the
Georgia Tech campus network for various DHCP lease times. The
existing lease time, 30 minutes, uses only 67.14% of the available
address space. As expected, longer lease times increase address
space utilization, particularly at peak times when many new clients
enter the network. A lease time of 240 minutes exhausts the avail-
able address space (a /20) at peak utilization; lease times of 90 and
150 minutes use about 80% and 90% of the address space, respec-
tively. A 90-minute lease period incurs a peak utilization of 81.54%
over a week and an average utilization of 79.47%.
6.2 Reducing DHCP Renewals and Expires
We measured the number of client expires saved as a result of
increasing the lease period from 30 minutes to 90 minutes over
5 days. Table 1 summarizes the savings in renewals and expires.
We found that an average of 8,398 expires occur per day with
30-minute lease times; about 6,483 expires occur with 90-minute
leases. On the Georgia Tech campus network, this implies that each
day there are about 1,915 instances where the clients could avoid
re-authentication if the lease period were 90 minutes, which would
improve for clients and decrease load on the authentication server.
Figure 5 shows example DHCP sessions for 9 clients on Febru-
ary 12, 2007 to help provide intuition for why longer lease times
can save re-authentications. The solid lines indicate the periods
when the client was active (as determined from DHCP renew re-
quests at 15-minute intervals). Clients 1,3,8 and 9 stand to benefit
as a result of an increase in the lease period. In a network like Geor-
gia Tech’s where each expired client has to re-authenticate itself af-
ter its DHCP lease expires, six of Client 9’s sessions are coalesced
into just two sessions; as a result, Client 9 must re-authenticate only
once, instead of five times.3
2This target utilization figure is a rule of thumb, but our results and analysis
techniques would apply generally for other target utilization figures as well.
3Figure 5 also shows an artifact of our estimation where clients appear to
have longer sessions with 30-minute lease times than with 90-minute lease
times. Consider Client 7. In the 30-minute case, this client renews at the
15th minute and 30th minute after its start of session. From this, we estimate
the client’s on-time to be 37.5 minutes. This on-time will not generate any
renewals in the replay log for the 90-minute case because the first expected
0 180 360 540 720 900 1080 1260 1440 1620 1800 1980 2160 2340 2520 2700 2880
Time (in minutes)
600
800
1000
1200
1400
1600
1800
2000
2200
2400
2600
2800
3000
3200
3400
3600
N
um
be
r o
f a
ct
iv
e
le
as
es
Exponential
Static: 90 min
Single Adj: 90-30
Original: 30 min
Figure 6: Address space utilization for two days for different strategies.
Strategy Consumed Expires Saved Renews Saved
Orig. (30 min) 2,750 (67.14%) 8,398 (———) 175,088 (———)
1-Adj. (90-30) 2,950 (72.02%) 7,469 (11.06%) 124,291 (29.01%)
Static (90 min) 3,340 (81.54%) 6,483 (22.80%) 50,920 (70.92%)
1-Adj. (150-30) 3,350 (81.78%) 6,415 (23.61%) 112,157 (35.94%)
Exponential 3,450 (84.22%) 6,163 (26.62%) 38,615 (77.94%)
Table 1: Total addresses consumed (and percent utilization of the /20
address space), the number of DHCP lease expires saved, and the num-
ber of renewals saved, for various lease setting strategies.
Longer lease times reduce renewal traffic both on the network
and on the DHCP server. We found that in the 90-minute case, an
average of only 50,920 renewal messages are seen per day, versus
175,088 per day with 30-minute lease times; a 90-minute lease pe-
riod thus saves up to 70.92% of DHCP renewal messages without
significantly increasing DHCP utilization above 80%.
Although it might seem natural to set the DHCP lease time to
about the time of a typical class period; however, in practice, usage
patterns are highly variable. In particular, class periods are variable,
wireless networks are deployed in many other parts of the campus
(e.g., common areas), and the campus network must support an
increasingly wider range of devices that use dynamic addressing
(e.g., gaming consoles, phones). As these devices place increasing
pressure on available address space, setting the appropriate lease
time for each client becomes increasingly important. In the next
section, we explore some strategies for setting the DHCP lease time
dynamically to account for this variability in usage patterns.
7. Dynamic Lease Time Optimization
In this section, we explore two strategies for dynamically adjust-
ing DHCP lease times for each client: single adaptation (whereby
a client receives a different lease time after the first REQUEST) and
exponential (whereby the lease time a client receives doubles each
time it renews a lease, up to a maximum possible lease time).
7.1 Single Adaptation
Our results from Section 4 show that client on-times are con-
centrated within certain time ranges: specifically, Figure 2 shows
that half of the client sessions are less than 75 minutes. Human be-
renewal in the replay is at 45 minutes. As such, we estimate the client’s on-
time in the 90-minute replay log to be 22.5 minutes. This artifact affects our
estimates of client session times but does not greatly affect our estimates for
the number of renewals and premature session expires.

havior can explain this statistic: Georgia Tech has class periods of
60-minute and 90-minute durations, and the majority of the DHCP
users are students with laptops. This aspect suggests that increasing
the lease time to 60 or 90 minutes might reduce renewals.
Accordingly, we propose a single adaptation strategy, which sets
a long initial lease and reduces the lease times for subsequent client
renewals. This strategy reduces lease times for the common case
but keeps the number of renewals and expires the same for the re-
maining cases. We attempted several single adaptation settings, two
of which are summarized in Table 1. The best strategy we found
was to set a lease period of 90 minutes for the initial REQUEST
and to reduce subsequent lease times to 30 minutes. The initial
lease period of 90 minutes also allows a user to have an off-time of
anywhere from 0 to 90 minutes, thus preventing expires for clients
that temporarily disconnect and re-attache to the network within a
short time period (e.g., a single class period).
Single adaptation has two main advantages. First, it requires
only minimal changes to the DHCP server, which need only dis-
tinguish DHCP Init-Reboot REQUEST messages from Renew-
ReInit REQUEST messages. This differentiation is permitted by
the DHCP specification and requires no additional state at the client
or server [2]. Second, this scheme offers a tradeoff between in-
creasing a lease time from a fixed lower value to a fixed upper value
(30 and 90 in our case), as shown in Figure 6.
As shown in Table 1, single adaptation with an initial lease time
of 90 minutes and subsequent lease times of 30 minutes increased
address space consumption by only 5% (as opposed to a 14% in-
crease with a static 90-minute lease) while still saving on 11% (929
instances) expires and 29% renewals each day. Unlike previous
work [1], our DHCP clients also included was collected from all
over the Georgia Tech network including areas where the expected
on-times of 60 minutes and 90 minutes may not be the common
case. We expect that single adaptation might perform even better if
applied only to clients using access points near classrooms.
7.2 Exponential Adaptation
The exponential adaptation strategy issues small leases to clients
when they first arrive (which serves clients with small on-times)
and doubles the lease every time the client does a renew (which
reduces the number of times that persistent clients need to renew
their leases). Because the scheme assumes a short lease time but
gracefully adapts for clients that remain active for long periods of
time, we expect that it will work well for networks whose usage
characteristics are unpredictable or otherwise not well known. We
experimented with an exponential strategy starting off with a lease
time of 30 minutes and doubling all the way up to 960 minutes.
We determined this upper bound by observing that for a single day,
there were only a few clients (about 1%) who had off-times of more
than 930 minutes. This strategy is motivated by the observation that
if a client has been active long enough to do a renew, then it could
be expected to stay on longer and deserves to be given a lease pe-
riod of double the previous value. This scheme does require that
the servers maintain state of the clients or that the clients report
their current or expected lease time in the options field of their RE-
QUEST, which is supported, though not required, by the protocol
specification.
Figure 6 shows that exponential adaptation increases address
space utilization during periods of low utilization but does not in-
crease peak address space utilization substantially over fixed 90-
minute lease times (presumably because it can also quickly reclaim
leases for clients that are more transient. Exponential adaptation
saves the most expires (26.62%) and renewals (77.94%), with the
address space utilization reaching a maximum of 84.22%. As op-
posed to static lease time settings, exponential adaptation reduces
client renewals by not forcing persistent clients to frequently renew.
Figure 6 shows that exponential adaptation can save renewals for
these persistent clients: The minimum address space consumption
never drops below 1,600 concurrent addresses, compared to about
800 more active leases than the minimum of any other scheme.
With exponential adaptation, any client with an on-time of more
than 465 minutes (about 8 hours) receives a lease of 16 hours,
thus allowing the client to return the next day without having to
re-authenticate.
8. Conclusion
Despite DHCP’s widespread usage and the importance of prop-
erly configuring DHCP client lease times, today’s network opera-
tors have little understanding of how to best optimize lease times
given an address pool and a group of dynamic users. This paper
takes a first step towards demystifying this process with three main
contributions: First, we study DHCP usage patterns on a campus-
wide network with a peak usage of more than 2,500 users and 1,000
access points. Second, we present a tool that uses existing DHCP
traces to analyze the effects of increasing lease times on DHCP
traffic, lease expiration, and address space utilization. This tool can
help operators evaluate the effects of alternate lease time settings
and strategies without having to experiment with these changes on
a running network. Finally, we use this tool to evaluate both the
effects of longer lease times and alternate strategies for adjusting
client-specific lease times. We find that dynamic lease time adjust-
ment strategies can significantly reduce the amount of DHCP traf-
fic and premature client session expirations without prohibitively
increasing peak address space utilization.
The techniques we have presented could apply to any DHCP net-
work. In future work, we would like to explore how alternative
lease time strategies could apply on other networks with dynamic
addressing that might have much different characteristics (e.g., ca-
ble modem users, users in developing regions).
Acknowledgments
We thank David Andersen for the suggesting the exponential lease
time strategy. This research was funded by NSF CAREER Award
CNS-0643974 and NSF grants CNS-0626950 and CNS-0721581.
REFERENCES
[1] V. Birk, J. Stroik, and S. Banerjee. Debugging DHCP
Performance. In Proc. Internet Measurement Conference,
Taormina, Italy, Oct. 2004.
[2] R. Droms. Dynamic Host Configuration Protocol, Mar. 1997.
RFC 2131.
[3] Georgia Tech Local Area Walkup/Wireless Network (LAWN).
http://www.lawn.gatech.edu/, 2007.
[4] R. Hutchins and E. W. Zegura. Measurements from a Campus
Wireless Network. In IEEE International Conference on
Communications, volume 5, pages 3161–3167, 2002.
[5] C. Perkins and K. Luo. Using DHCP With Computers that
Move. Wireless Networks, 1(3):341–353, Sept. 1995.