Bandwidth management VMs live migration in wireless fog computing for 5G networks

1 MB·

by  Danilo Amendola

Bandwidth management VMs live migration
in wireless fog computing for 5G networks
Danilo Amendola Nicola Cordeschi Enzo Baccarelli
DIET, Sapienza University of Rome DIET, Sapienza University of Rome DIET, Sapienza University of Rome
[email protected] [email protected] [email protected]
Abstract—Live virtual machine migration aims at enabling This paper is organized as follows. Section II gives a
the dynamic balanced use of the networking/computing physical brief description of the live migration problem. Section III
resourcesofvirtualizeddata-centers,sotoleadtoreducedenergy
introduces our bandwidth manager, formulation and solution
consumption. Here, we analytically characterize, prototype in
of the non-convex optimization problem. Section IV shows
software and test an optimal bandwidth manager for live mi-
7 gration of VMs in wireless channel. In this paper we present experimental work and tests. Finally, we conclude our work
1 theoptimaltunable-complexitybandwidthmanager(TCBM)for in Section V.
0 the QoS live migration of VMs under a wireless channel from
2 smartphone to access point. The goal is the minimization of II. THETACKLEDPROBLEM:LIVEMIGRATION
themigration-inducedcommunicationenergyunderservicelevel
n Virtualization is a viral technology in the data center and
agreement(SLA)-inducedhard constrainsonthetotalmigration
a hardware efficient utilization, its benefit is well recognized in
time, downtime and overall available bandwidth.
J
Keywords- Bandwidthmanagement;Optimization;Quality a large number of applications. Virtualization [8] is rapidly
2
of Service; Energy-saving; Live migration. evolving and live migration is a core function to replace
2
running VMs seamlessly across distinct physical devices [9].

] I. INTRODUCTIONANDRELATEDWORK In recent years considerable interest has been pointed out
I Mobilecloudcomputing(MCC)emerginginthecontextof onVMlivemigrationfordatacentermigration[9]andcluster
N
5G has the potential to overcome resource limitation in the computing.
.
s mobile devices (appear as a bottleneck in 5G applications), Efficient VM live migration is an attractive function in vir-
c
which enables many resource-intensive services for mobile tualizedsystemscausethisisessentialtoenableconsolidation
[
users with the support of mobile big data delivery and cloud- techniques oriented to save energy consumption. Representa-

1 assisted computing [1]. tive technologies for VM live migration are XenMotion [9]
v
In 5G a fundamental issue is to provide services with low and VMware VMotion, both of them implemented as a built-
8
7 latency Fog computing (FC), also know as edge computing, in tool in their virtualized platforms. There are also other
1 canaddressthoseproblemsbyprovidingelasticresourcesand studies about VM migration in which the problem of where
6 servicestoendusersattheedgeofthenetwork.Thedifference and when a VM should be migrated to improve the system
0 between fog computing and cloud computing (CC) is that CC performancesisconsidered.Butnoneofthemareconsidering
.
1 focuses on providing resources located in the core network, the issue of how to improve the communication performance
0 whileFCfocusesonresourcesdistributedintheedgenetwork. with bandwidth optimization for migration when time and
7
In this context a plethora of frameworks and models (ori- place of migration are decided.
1
entedtotheMCC)areproposed,toprovidehighperformance ThenVMlivemigrationtechnologiesareveryeffectivetool
:
v computation system on mobile devices. We briefly present in to enable data-center management and save energy consump-
i
X the follows some of these solutions: tion. During the live migration, physical memory image is
r • CloneCloud [2], [3]: is a system that has the ability to transferred across the network to the new destination, while
a automaticallytransformmobiledeviceapplicationinsuch the source VM continue to run until the last bit will be
a way that they can run into the cloud; received with success. We treated this issue in our previous
• VOLARE [4]: is a middelware-based solution which work [10], we considered the intra data-center channel opti-
allowscontext-awareadaptivecloudservicediscoveryfor mization bandwidth problem. Hence, here we investigate live
the mobile devices. virtualmachinemigrationbandwidthoptimizationonwireless
• Cuckoo [5]: is a computational offloading framework for channel. Besides, these works [11], [12], [13], [14] are useful
mobile devices; to understand our approach.
• Cloudlet [6]: is a set of widely dispersed and decentral- In literature there are four main techniques for VM mi-
ized Internet infrastructure components, with non-trivial gration, namely, stop-and-copy migration (SaCM), pre-copy
characteristic to make available for the nearby mobile migration (PeCM), post-copy migration (PoCM) and hybrid
devices computing resource and storage resources; migration (HyBM). They trade-off the total migration time
• MAUI [7]: is a system that is able to minimize the and downtime. Here, to be concise, we omitted a complete
energyduetotheVMmigrationbymeansoffine-grained overview of main techniques for VM live migration. To
offloading. understand how it works you can refer to our work [10]. In
the following we use the pre-copy live migration technique, andT playtheroleofconstantparameters,inthesequel,we
DT
as in [10]. Our approach may be applied to all the mentioned focusontheevaluationofthe(alreadydefined)stop-and-copy
techniques. time T and the resulting memory migration time T ,
SC MMT
Considering the related work, at this time there are not which is defined as in:
worksconsideringthebandwidthmanagementduringtheVMs
T ≡ T (R) (cid:44) T (R)+T (R). (1)
livemigrationforwirelesschannel.Ourpreviouswork[10]is MMT MMT IP SC
the first which considers the bandwidth management in wired
Table I reports the definitions of the key parameters used in
network environment. In that work we presented a bandwidth
the paper. Since the PeCM technique performs the iterative
manager atop an intra-data-center wired test-bed comparing
pre-copy of dirtied memory bits over consecutive rounds, let
performances with most relevant VMs live technologies.
V (Mb) and T (s), i = 0,...,(I +1), be the volume
i i MAX
As we described in [10], this approach is capable to effec-
of the migrated data and the time duration of the ith round,
tively filter out transient fluctuations of the average resource
respectively.Bydefinition,V andT arethememorysizeM
0 0 0
utilization and avoid needless migrations [15].
(Mb) of the migrating VM and the time needed for migrating
it during the 0th round, respectively.
III. TUNABLECOMPLEXITYBANDWIDTHMANAGEMENT
DEFINITIONANDBASICPROPERTIES
TABLE I: Main taxonomy of the paper.
In this section we introduce the tunable complexity band-
width management (TCBM). Let I be the number of Symbol Meaning/Role
max
performed pre-copy rounds. IMAX Numberofmigrationpre−copyrounds
A primary goal of our work is to formal define a model i Roundindex,i=0,...,(IMAX+1)
w(Mb/s) MemorydirtyrateofthemigratedVM
overview of how live migration works. Most important vari- Rˆi (Mb/s) Migrationbandwidthatithround
ables are total migration time and downtime. From a formal P(Ri)(W) CommunicationpoweratthemigrationbandwidthRi
pointofview,thetotalmigrationtimeT (s)istheoverall Rˆ (Mb/s) Maximumavailablemigrationbandwidth
TOT
duration:T (cid:44)T +T +T +T +T +T ,of M0 (Mb) MemorysizeofthemigratedVM
the six stagTeOsT(as wePMcan seReEin FiIgP. 1), SwChile tCheMdowAntTime: ETOT (J) Totalconsumedcommunicationenergy
TDT (cid:44) TSC + TCM + TAT, is the time required for the ∆∆MSCM(Ts)(s) MMaaxxiimmuummttoolleerraatteeddmstoepm−oaryndm−igcoraptyiotnimtieme
execution of the last three stages. From a practical point of β Migrationspeed−upfactor
n Integer−valuediterationindex
view, T is the period when the states of the source and
TOT
destination servers must be synchronized, which may also
affect the reliability of the migration process, while T is Now we formalize the afforded tunable-complexity band-
DT
theperiodinwhichthemigratingVMishaltedandtheclients width manager. In addition to R and R we have Q,
0 IMAX+1
experience a service outage [16]. which is the number of updated rates. Then we updated Q
out of I rates of the pre-copy rounds evenly spaced by
MAX
I
S (cid:44) MAX over the round-index set {1,2,3,...,I }.
Q MAX
For this purpose, we perform the partition of the round
indexset{1,2,3,...,I }intoQnotoverlappingcontiguous
MAX
subsets of size S.
Fig.2:Referenceframeworkforthetunable-complexityband-
width manager. Case of I = 6 , Q = 3. The rates to be
MAX
Fig. 1: Pre-copy live migration stages (six stages). uploaded are: R ,R ,R ,R and R . The rates to be held
0 1 3 5 7
are: R ≡R ; R ≡R ; R ≡R .
2 1 4 3 6 5
Let R (Mb/s) be the transmission rate used during the
i
thirdandfourthstagesattheithroundformigratingtheVM, The fist rate R , j =0,...,(Q−1) of each subset is
jS+1
that is, the migration bandwidth. However, we present a first updated, while the remaining (S−1) rates are set to R ,
jS+1
formulationoftheproblemconsideringaconstantrateateach that is R ≡R , for i=jS+2, jS+3, ..., (j+1)S.
i jS+1
round, R = R∀i. Since, by definition, only T and T Fig. 2 illustrates the framework of the updated/held migra-
i IP SC
dependonR,whilealltheremainingmigrationtimesinT tionratesforthedummycaseofI =6andQ=3.Inthis
TOT MAX
case,R ,R ,R ,R andR aretheQ+2=5migrationrates default setting; R is: 0.9 × 2(Mb/s) for 3G cellular;
0 1 3 5 7 MAX
to be updated, while R ,R and R are the (I −Q)=3 0.9×11(Mb/s) for IEEE 802.11b; and 0.9×50(Mb/s) for
2 4 6 MAX
migration rates which are not updated and, by definition, they 4G-LTE, where R (Mb/s) is the maximum throughput
MAX
equate: R ≡R ; R ≡R ; R ≡R . at the Transport Layer. E is: 3.25(J) for 3G cellular;
2 1 4 3 6 5 SETUP
5.9(J)forIEEE802.11.b;5.1(J)for4G-LTE.whereE
SETUP
A. Formulation of the non-convex optimization problem to be
is the static (e.g., rate independent) part of the overall energy
solved by the TCBM
consumption of the considered connection.
The TCBM is the solution of the following non-convex Alltestshavebeen carriedoutinthreedifferentapplication
optimization problem, which could be solved as an instance scenarios, i.e., the scenario in which the smartphone migrates
of geometric problem (solution is omitted here for briefness, totheaccesspointby3g;thescenarioinwhichthesmartphone
see [10] for details): migrateswiththeuseofthe4G;andfinallythescenariowhere
migration is performed by making use of WiFi.
min E (2)
TOT
{R0,RjS+1,j=0,1,...,(Q−1);RIMAX+1} A. The benchmark Xen bandwidth management
s.t. The currently implemented Xen hypervisor adopts a pre-
(cid:26)(cid:18) (cid:19) (cid:27)
Ψ (cid:44)θ 1 T −1 ≤0; (3) copy heuristic bandwidth management policy, which operates
1 ∆TM TM on a best effort basis, while attempting to shorten the final
(cid:18) (cid:19) stop-and-copy time [21], [22]. The rationale behind this Xen
1
Ψ2 (cid:44) ∆DTTDT −1≤0; (4) policy is that, in principle, the stop-and-copy time may be
reduced by monotonically increasing the migration bandwidth
(cid:26) (cid:27)
Ψ (cid:44)θ βwR−1−1 ≤0, over consecutive rounds [22]. For this purpose, the Xen
3 i (5) hypervisor uses pre-assigned minimum: RXEN (Mb/s), and
MIN
fori=0; i=jS+1; j =0,...,(Q−1); maximum: RXEN (Mb/s) bandwidth thresholds, in order
MAX
to bound the migration bandwidth during the pre-copy stage
Ri ≤R(cid:98), (6) (see Section 5.3 of [22]). Specifically, the Xen migration
fori=0; i=jS+1; j =0, ...,(Q−1); i=IMAX +1; bandwidth RXEN equates: RXEN (Mb/s) at round#0, and,
MIN
Four constraints are considered in the formulation of the then, it increases in each subsequent round by a constant
TCBM, which capture, in turn, the metrics currently adopted term: ∆RXEN (Mb/s), so to reach the maximum value:
for measuring the performance of live migration techniques RXEN = RXEN at the last round: round#(I + 1) (see
MAX MAX
[17], [18]. The first two constraints upper limit the tolerated Section 5.3 of [22]). In the carried out field trials, we have
total migration time and downtime. Constrain (5) account the implemented this benchmark policy by setting:
ratio of the volumes of data migrated over two consecutive
rounds falls below a predefined speed-factor β > 1. Finally, ∆RXEN =(RXEN − w)/(IXEN +1), (7)
MAX MAX
constrain (6) upper limit the maximum available rate. Fur-
and
thermore, the θ parameter in (3) accounts for the fact that,
by definition, the total migration and stop-and-copy times RXEN =w + i∆RXEN, i=0,...,(IXEN +1). (8)
i MAX
coincide under the SaCM and PoCM techniques.
We point out that, on the basis of the (recent) surveys in
IV. EXPERIMENTALWORKANDTESTS [17], Chapter 3 of [21] and Chapter 17 of [23], this is the
In order to actually test and compare the performance of only bandwidth management policy currently considered by
the proposed bandwidth manager, we have implemented an bothacademyandindustryforVMmigration.Thisisalsothe
experimental wireless test-bed. bandwidth policy currently implemented by Xen, KVM and
Belowwediscusssomeexperimentsthatshowthegoodness VMware commercial hypervisors [21].
of our TCBM, comparing with the results obtained from Xen
B. Tracking capabilities under contention phenomena
andthemethodBMOP(BandwidthManagementOptimization
Problem, see [10] for implementation) in which, unlike in our Real-worldapplicationsmayvarytheproducedtrafficsover
software, the initial rate, is held for the entire duration of the the time [24] and, then, it may be of interest to test how
VM migration . the proposed bandwidth manager reacts when the workload
Of practical interest, specifically, the reported data refer to offered by the migrating VM changes unexpectedly.
the average parameters of typical wireless IEEE 802.11b, 3G- As pointed out in [17], memory contention phenomena
UTRAN and 4G-LTE connections. We anticipated that the and/or network congestions may produce abrupt (typically,
reported data are in agreement with [19] for 3G-UTRAN and unpredictable) time-variations of the parameters w and or K
0
[20] for 4G-LTE. Hence, in order to evaluate the tracking capabilities of the
After noting that I˜ refers to our optimized setting of proposedadaptivebandwidthmanageranditssensitivitytothe
MAX
the allowed pre-copy rounds, typically values for the tested parametersa inFig.3,wereportthemeasuredbehaviors
MAX
VMs are: 1 ≤ I˜ ≤ 29, where I˜ =29 is the Xen’s oftheenergysequence:{E∗(n), n≥0}when,duetomemory
MAX MAX TOT
700 Energyconsumption3G:time-varyingWavg=(0.8,1.5,0.8) 10000Energyconsumption4G:time-varyingWavg=(11.25;24;11.25) 1200 EnergyconsumptionWiFi:time-varyingWavg=(4;8;4)
aMAX=0.05 9000 aMAX=0.05 aMAX=0.05
600 aMAX=0.01 8000 aMAX=0.01 1000 aMAX=0.01
500 aMAX=0.005 7000 aMAX=0.005 800 aMAX=0.005
Energy(J)400 Energy(J)456000000000 Energy(J)600
300 3000 400
2000
200 200
1000
1000 10 20 30 Itera40tionind50exn 60 70 80 90 00 10 20 30 Itera40tionind50exn 60 70 80 90 00 10 20 30 Itera40tionind50exn 60 70 80 90
(a)w=[0.8,1.5,0.8] (b)w=[11.25,24,11.25] (c)w=[4,8,4]
Fig. 3: Time evolutions (in the n index) of the energy consumption of the proposed bandwidth manager, case of time-
varying w, at: (a) R(cid:98) = 1.8 (Mb/s), M0 = 256 (Mb), β = 2, ∆TM = 1460 (s), ∆DT = 0.14 (s), for 3G scenario;
(b) R(cid:98) = 45 (Mb/s), M0 = 256 (Mb), β = 2.33, ∆TM = 58.6 (s), ∆DT = 5.61 × 10−3 (s), for 4G scenario; (c)
R(cid:98)=9.9(Mb/s), M0 =256(Mb), β =2.33, ∆TM =266(s), ∆DT =2.55×10−2(s), for WiFi scenario.
111468000000 Energyconsumption3G:time-varyingK0=aaaMMM(0AAA.XXX18,===1.8000,0...000.151085) 1102000000Energyconsumption4G:time-varyingK0=aaa(MMM0AAA.0XXX9,===0.9000,...0000.501059) 111468000000EnergyconsumptionWiFi:time-varyingK0=aaaMMM(0AAA.XXX05===,0.0005...,0000510.505)
Energy(J)11680200000000 Energy(J)68000000 Energy(J)11680200000000
4000
400 400
200 2000 200
00 10 20 30Itera4t0ionin5d0exn60 70 80 90 00 10 20 30Itera40tionin5d0exn60 70 80 90 00 10 20 30 Itera40tionin5d0exn60 70 80 90
(a)K0=[0.18,18,0.18] (b)K0=[0.09,0.9,0.09] (c)K0=[0.05,0.5,0.05]
Fig. 4: Time evolutions (in the n index) of the energy consumption of the proposed bandwidth manager, case of time-
varying K0, at: (a) R(cid:98) = 1.8 (Mb/s), M0 = 256 (Mb), β = 2, ∆TM = 1460 (s), ∆DT = 0.14 (s), for 3G scenario;
(b) R(cid:98) = 45 (Mb/s), M0 = 256 (Mb), β = 2.33, ∆TM = 58.6 (s), ∆DT = 5.61 × 10−3 (s), for 4G scenario; (c)
R(cid:98)=9.9(Mb/s), M0 =256(Mb), β =2.33, ∆TM =266(s), ∆DT =2.55×10−2(s), for WiFi scenario.
contentionphenomena,thememorydirtyratewoftherunning by the software to return to the equilibrium state.
memtester application abruptly varies.
For this reason we prefer to work with a high, over
MAX
An examination to the plots of Fig. 3 and Fig. 4 supports the interval [0.5 , 0.05], in such a way that (in a maximum of
three main conclusions. six or seven iterations), the software reacts well to variations
of w.
• First, according to the fact that the energy function
Overall, from the outset, we conclude that the proposed
increasesforincreasingw and/orK ,alltheplotsofFig.
0 adaptive bandwidth manager is robust with respect to the
3 and Fig. 4 scale up at n = 30 and, then, scale down at
actual tuning of a , at least for values of a ranging
n = 60. MAX MAX
over the the interval [0.5 , 0.05], in order to exhibits the
• Second, the proposed bandwidth manager quickly reacts
best trade-off among the contrasting requirements of short
to abrupt unpredicted time variations of the migrating
transient-states and stable steady-states.
application and/or underlying network connections.
• Third,whilevirtuallyindistinguishableplotsareobtained
C. Comparative energy tests under random migration order-
for a ranging over the interval [5×10−2,5×10−3]
MAX ing and synthetic workload
incaseoftime-varyingK (see.Fig.4),thesameresults
0
is not obtained in case of time-varying w (see. Fig. 4). The benchmark bandwidth management policy of the Xen
This phenomenon is due to the fact that while K is a hypervisor does not guarantee, by design, minimum energy
0
multiplicativeconstantintheformulaoftheenergy,w,in consumptions and does not enforce QoS constraints on the
addition to a profound impact on energy, causes that our resulting memory migration and stop-and-copy times. Fur-
TCBMusesmoreiterationstogofromtransient-statesto thermore, differently from I˜ , the maximum number of
MAX
thesteady-states.Precisely,itisshowedthat,thedecrease allowed rounds: IXEN is fixed by the Xen hypervisor in an
MAX
ofa increasesthenumberofiterationsthatareused application-oblivious way (typically, IXEN ≤ 29; see [25],
MAX MAX
TABLE II: Scenario 4G with M =256(Mb); α=2; K =
[21]). Hence, in order to carry out fair energy comparisons, 0 0
(cid:18) (cid:19)
w
in the carried out field trials, we proceed as follows: 0.09; ESETUP = 5.1(J); (a) = 0.33 and R(cid:98) = 0.33×
i) set IXEN and RXEN; (cid:18) R(cid:98)(cid:19)
MAX MAX w
ii) measure the resulting Xen energy consumption EXEN, RXEN = 14.85(Mb/s); (b) = 0.11 and R(cid:98) = 0.11×
speed-up factor βXEN, total migration time TTXOETN, MAX R(cid:98)
TM RXEN =4.95(Mb/s);
downtime TXEN; MAX
DT
iii) enforce R(cid:98) ≡RMXEANX, together with the QoS constraints: IMXEANX 6 14 25
∆ ≡TXEN, ∆ ≡TXEN, and β ≡βXEN;
iv) meTaMsure thTeMresultinDgTenergDyTconsumption E∗ of the TDXTEN =∆DT(s) 0.103 5.42×10−4 4.03×10−7
TOT
proposed bandwidth manager at IMAX =I˜MAX. TTXMEN =∆TM(s) 46.9 65.2 83.6
The memtester (see in [10]) is the application considered
β 1.87 1.95 1.98
in this section and the implemented migration ordering of the
dirtied memory pages is the random one. ETXOETN(J) 1880 2150 2470
Thenumericalresultsmeasuredthroughacampaignoftrials ELIV MIG(J) 1550 1550 1550
TOT
developed for the three considered scenarios (3G, 4G and
WiFi) are partially omitted cause the lack of space. We show Q 1 1 1
only the table data referred top the 4G scenario.
ETCBM(J) 1366 1373 1373
TOT
ThesetableshowtheenergyvaluesobtainedusingXen,the
bandwidth management policy developed in the paper [10], En.savevs.XEN(%) 27.3 36.1 44.4
and the Tunable-complexity bandwidth manager.
En.savevs.LIV MIG(%) 11.8 11.4 11.4
An examination of the results of data leads to two main
conclusion. First, in all the carried out field trials the percent (a)
energy saving:
• (1−(ET∗OT/ETXOETN))% of the proposed bandwidth man- IMXEANX 6 14 25
ager over the Xen one is between 3% (minimum value TDXTEN =∆DT(s) 5.9×10−4 4.07×10−9 3.4×10−16
o4f4.e4n%erg(myasxaviminugm) fvoarl(uwe/oR(cid:98)f)en=er0g.y11saavnidngI)MfoArX(w=/2R(cid:98)5), t=o TTXMEN =∆TM(s) 84.9 110 137
0.33 and I =25 (see Table II(a)); β 4.47 4.78 4.89
MAX
• (1 − (ET∗OT/ETLOIVTMIG))% of the proposed bandwidth EXEN(J) 632 624 602
manager over the BMOP (Bandwidth Management Op- TOT
timization Problem, see paper [10]) is between 11.2% ELIV MIG(J) 1170 1170 1170
TOT
(minimum value of energy saving) for (w/R(cid:98)) = 0.33
Q 1 1 1
and I = 6, to 54.5% (maximum value of energy
MAX
saving) for (w/R(cid:98)) = 0.11 and IMAX = 6 (see Table ETTOCTBM(J) 531.7 545.25 541.8
II(b)).
En.savevs.XEN(%) 15.8 12.6 10
In all scenarios, TCBM appears to be the best one from the
pointofviewofenergysaving.Thesenoticeableenergygains En.savevs.LIV MIG(%) 54.5 53.4 53.7
supporttheconclusionthatthebandwidthmanagementpolicy
(b)
developed in this paper is the optimal one, and, by design, it
minimizes the migration-induced energy consumption.
Second, the values of the measured energy gains mainly
Inalltheexperiments,R waschosenequaltothevalue
depend on the considered ratio: (w/R(cid:98)). In particular, in these MAX
of R of 3G (which turns out to be smaller, than those in
tests only values of (w/R(cid:98)) ≤ 0.33 are considered, because, MAX
the 4G and WiFi), in such a way to have the comparisons in
if and only if this constraint is satisfies, the Xen (heuristic)
a consistent manner.
bandwidth management policy presents decreasing values of
energy for increasing values of I . Hence, under this The Figures 5 show the results of the tests.
MAX
condition, it make sense to compare our bandwidth manager An examination of the bar plots of Fig.5 leads to two main
with Xen and BMOP. conclusion.First,sincethedirtyrateincreasesbypassingfrom
In the carried out tests, is reported that, while the TCBM the (read-intensive) bzip2 program to the (write-intensive)
in each scenario presents a constant gain with respect to the memcached one, the corresponding energy consumptions also
optimization method described in [10], from the comparison exhibit increasing trends under both the Xen, LIV-MIG [10]
withXencomesoutthatthepercentageofenergysavingtends and proposed bandwidth managers. Second, in all cases, the
to decrease (for increase of IMAX) when the ratio (w/R(cid:98)) < energy consumption relating to the migration by applying
0.33; on the contrary the percentage of energy saving tends to our method appears to be lower than both Xen and LIV-
increase when the ratio (w/R(cid:98))=0.33. MIG manager. In particular, the percent energy savings of the
3GScenario 4GScenario WiFiScenario
250 250 250
XENmanager 228(J) XENmanager 228(J) XENmanager 228(J)
LIV-MIGmanager LIV-MIGmanager LIV-MIGmanager
200 TCBMmanager 196(J) 200 TCBMmanager 196(J) 200 TCBMmanager 196(J)
Energy(J)110500 100(J) 107(J) 128(J)113.73(J) 135(J)126(J) Energy(J)110500 100(J) 107(J) 128(J) 135(J) Energy(J)110500 100(J) 107(J) 128(J) 135(J)
71.88(J) 66.48(J)
60.34(J)
50 50 39.41(J) 50 36.59(J) 40(J)
24.96(J)
0 0 0
bzip2 mcf memcached bzip2 mcf memcached bzip2 mcf memcached
(a)3GScenario (b)4GScenario (c)WiFiScenario
Fig. 5: Energy consumptions obtained by bzip2, mcf and memcached in : (a) 3G scenario; (b) 4G scenario; (c) WiFi scenario.
proposed manager over the Xen and the LIV-MIG under the [7] E. Cuervo, A. Balasubramanian, D.-k. Cho, A. Wolman, S. Saroiu,
bzip2, mcf and memcached,for eachapplicationscenariosare R. Chandra,and P.Bahl, “Maui: makingsmartphones lastlonger with
code offload,” in Proceedings of the 8th international conference on
reported in Table II and Table III.
Mobilesystems,applications,andservices. ACM,2010,pp.49–62.
[8] P.Barham,B.Dragovic,K.Fraser,S.Hand,T.Harris,A.Ho,R.Neuge-
TABLE III: Percent energy savings of the TCBM manager bauer,I.Pratt,andA.Warfield,“Xenandtheartofvirtualization,”ACM
over Xen and LIV-MIG managers. SIGOPSOperatingSystemsReview,vol.37,no.5,pp.164–177,2003.
[9] C.Clark,K.Fraser,S.Hand,J.G.Hansen,E.Jul,C.Limpach,I.Pratt,
Parameter bzip2 mcf memc. and A. Warfield, “Live migration of virtual machines,” in Proceedings
ofthe2ndconferenceonSymposiumonNetworkedSystemsDesign&
3G Energy saving resp. Xen(%) 28.1 41.92 44.74 Implementation,Boston,USA,2005,pp.273–286.
En. saving resp. LIV-MIG(%) 32.8 11.15 6.67 [10] E.Baccarelli,D.Amendola,andN.Cordeschi,“Minimum-energyband-
widthmanagementforqoslivemigrationofvirtualmachines,”Computer
Energy saving resp. Xen(%) 60.5 69.21 70.84
4G Networks, vol. 93, Part 1, pp. 1 – 22, 2015. [Online]. Available:
En. saving respect LIV-MIG(%) 63.17 52.86 50.76
http://www.sciencedirect.com/science/article/pii/S1389128615003655
Energy saving resp. Xen(%) 75.04 81.33 82.46 [11] E.Baccarelli,M.Biagi,R.Bruno,M.Conti,andE.Gregori,Broadband
WiFi En. saving resp. LIV-MIG(%) 76.67 71.41 70.37 wirelessaccessnetworks:aroadmaponemergingtrendsandstandards.
Wiley,2005.
[12] N. Cordeschi, T. Patriarca, and E. Baccarelli, “Stochastic traffic engi-
This confirms the trend of the previous Section IV-C about neering for real-time applications over wireless networks,” Journal of
NetworkandComputerApplications,vol.35,no.2,pp.681–694,2012.
thelargeenergy-gainsofferedbytheproposedmanagerunder
[13] N.Cordeschi,D.Amendola,andE.Baccarelli,“Hardandsoftoptimal
write-intensive applications.
resource allocation for primary and secondary users in infrastructure
vehicular networks,” in 2015 12th Annual IEEE Consumer Communi-
V. CONCLUSION cationsandNetworkingConference(CCNC),Jan2015,pp.708–713.
[14] ——, “Resource-management for vehicular real-time application
In this paper we presented a novel approach for bandwidth under hard reliability constraints,” in Proceedings of the 2014
IEEE/ACM 18th International Symposium on Distributed Simulation
managementinlivemigrationvirtualmachineinwirelesscon- andRealTimeApplications,ser.DS-RT’14. Washington,DC,USA:
text. Our results show a significant improvement with respect IEEE Computer Society, 2014, pp. 219–226. [Online]. Available:
tothecurrentlyusedapproachinmostrelevantimplementation http://dx.doi.org/10.1109/DS-RT.2014.34
[15] T.Wood,P.J.Shenoy,A.Venkataramani,andM.S.Yousif,“Black-box
architecture for live virtual machines.
andgray-boxstrategiesforvirtualmachinemigration,”inNSDI,vol.7,
Cambridge,USA,2007,pp.229–242.
REFERENCES [16] F. Xu, F. Liu, H. Jin, and A. V. Vasilakos, “Managing performance
overheadofvirtualmachinesincloudcomputing:Asurvey,stateofthe
art,andfuturedirections,”ProceedingsoftheIEEE,vol.102,no.1,pp.
[1] M.Chen,Y.Zhang,Y.Li,S.Mao,andV.Leung,“Emc:emotion-aware
11–31,2014.
mobilecloudcomputingin5g,”Network,IEEE,vol.29,no.2,pp.32–
38,2015. [17] ——, “Managing performance overhead of virtual machines in cloud
[2] B.-G. Chun, S. Ihm, P. Maniatis, M. Naik, and A. Patti, “Clonecloud: computing:asurvey,stateoftheart,andfuturedirections,”Proceedings
elasticexecutionbetweenmobiledeviceandcloud,”inProceedingsof oftheIEEE,vol.102,no.1,pp.11–31,2014.
thesixthconferenceonComputersystems. ACM,2011,pp.301–314. [18] Y. Wu and M. Zhao, “Performance modeling of virtual machine live
[3] M. Schu¨ring, “Mobile cloud computing–open issues and solutions,” in migration,” in Cloud Computing (CLOUD), 2011 IEEE International
15thTwenteStudentConferenceonIT,Enschede,TheNetherlands,2011. Conferenceon. IEEE,2011,pp.492–499.
[4] P. Papakos, L. Capra, and D. S. Rosenblum, “Volare: context-aware [19] G. P. Perrucci, F. H. Fitzek, and J. Widmer, “Survey on energy con-
adaptivecloudservicediscoveryformobilesystems,”inProceedingsof sumptionentitiesonthesmartphoneplatform,”inVehicularTechnology
the9thInternationalWorkshoponAdaptiveandReflectiveMiddleware. Conference(VTCSpring),2011IEEE73rd. IEEE,2011,pp.1–6.
ACM,2010,pp.32–38. [20] J. Huang, F. Qian, A. Gerber, Z. M. Mao, S. Sen, and O. Spatscheck,
[5] R.Kemp,N.Palmer,T.Kielmann,andH.Bal,“Cuckoo:acomputation “A close examination of performance and power characteristics of 4g
offloadingframeworkforsmartphones,”inMobileComputing,Applica- lte networks,” in Proceedings of the 10th international conference on
tions,andServices. Springer,2010,pp.59–79. Mobilesystems,applications,andservices. ACM,2012,pp.225–238.
[6] M.Satyanarayanan,P.Bahl,R.Caceres,andN.Davies,“Thecasefor [21] K. Hwang, J. Dongarra, and G. C. Fox, Distributed and cloud com-
vm-basedcloudletsinmobilecomputing,”PervasiveComputing,IEEE, puting: from parallel processing to the internet of things. Morgan
vol.8,no.4,pp.14–23,2009. Kaufmann,2013.
[22] C.Clark,K.Fraser,S.Hand,J.G.Hansen,E.Jul,C.Limpach,I.Pratt,
and A. Warfield, “Live migration of virtual machines,” in Proceedings
ofthe2ndconferenceonSymposiumonNetworkedSystemsDesign&
Implementation-Volume2. USENIXAssociation,2005,pp.273–286.
[23] H. T. Mouftah, Communication Infrastructures for Cloud Computing.
IGIGlobal,2013.
[24] N. Cordeschi, M. Shojafar, D. Amendola, and E. Baccarelli, “Energy-
efficientadaptivenetworkeddatacentersfortheqossupportofreal-time
applications,”TheJournalofSupercomputing,vol.71,no.2,pp.448–
478,2015.
[25] D. Chisnall, The definitive guide to the xen hypervisor. Pearson
Education,2008.