Temporal Isolation Among LTE/5G Network Functions by Real-time

Scheduling

∗

Tommaso Cucinotta

, Mauro Marinoni

, Alessandra Melani

, Andrea Parri

and Carlo Vitucci

Scuola Superiore Sant’Anna, Pisa, Italy

Ericsson, Stockholm, Sweden

Keywords:

Real-time Scheduling, Virtual Radio Access Network, Temporal Isolation, Quality of Service Control.

Abstract:

Radio access networks for future LTE/5G scenarios need to be designed so as to satisfy increasingly stringent

requirements in terms of overall capacity, individual user performance, ﬂexibility and power efﬁciency. This is

triggering a major shift in the Telcom industry from statically sized, physically provisioned network appliances

towards the use of virtualized network functions that can be elastically deployed within a ﬂexible private cloud

of network operators. However, a major issue in delivering strong QoS levels is the one to keep in check the

temporal interferences among co-located services, as they compete in accessing shared physical resources.

In this paper, this problem is tackled by proposing a solution making use of a real-time scheduler with strong

temporal isolation guarantees at the OS/kernel level. This allows for the development of a mathematical

model linking major parameters of the system conﬁguration and input trafﬁc characterization with the achieved

performance and response-time probabilistic distribution. The model is veriﬁed through extensive experiments

made on Linux on a synthetic benchmark tuned according to data from a real LTE packet processing scenario.

1 INTRODUCTION

Radio access networks (RANs) are undergoing a

steep evolution towards higher overall capacity and

coverage across the territory and enhanced perfor-

mance for individual mobile devices, both in terms of

throughput and round-trip latency towards cloud com-

puting services (L

opez-P

erez et al., 2015; Clo, 2015;

Cis, 2016; Fiorani et al., 2015). This, coupled with an

increasing need for power-efﬁcient and adaptive solu-

tions supporting highly variable trafﬁc conditions, re-

sults in a strong demand for ﬂexible network architec-

tures, designed applying principles from distributed,

cloud and service-oriented computing, thanks also to

the progressive convergence towards IP-based net-

works, as witnessed by LTE.

Traditional deployments suffer of a great under-

utilization of resources throughout the majority of

(off-peak) operations, leading to drawbacks in terms

of energy consumption, as well as unsustainable

OPEX. As opposed to this, the novel paradigm of Net-

work Function Virtualization (NFV) (NFV Industry

∗

This work was partially funded by Ericsson and has

been partially supported by the RETINA Eurostars Project

E10171. The authors want to thank Freescale for their sup-

port with the reference platform.

Specif. Group, 2012) is spreading, where IP-based

network functions are virtualized into software com-

ponents which are deployed, consolidated and scaled

across a ﬂexible infrastructure, managed according to

a private cloud deployment model (Armbrust et al.,

2009; Hawilo et al., 2014).

Design of Telecom applications needs tradition-

ally to focus not only on functional correctness, but

also on non-functional properties such as availability

and performance. The latter ones have been facili-

tated by classical designs based on having separate

physical appliances dedicated to different network

functions. However, the efﬁciency of new NFV-based

designs relies on aggressive consolidation strategies

where various possibly heterogeneous functions (e.g.,

GSM, LTE, IMS, security-related, billing, monitor-

ing) can be hosted on the same system, leading to

increased temporal interferences among them due to

competing accesses to shared physical resources.

Therefore, one of the key success factors in new

designs is the degree at which the various co-located

functions can actually be temporally isolated from

each other. The traditional way to control the recip-

rocal interference among co-deployed software com-

ponents is by partitioning the available physical re-

sources. For example, a network function can be de-

368

Cucinotta T., Marinoni M., Melani A., Parri A. and Vitucci C.

Temporal Isolation Among LTE/5G Network Functions by Real-time Scheduling.

DOI: 10.5220/0006246703680375

In Proceedings of the 7th International Conference on Cloud Computing and Services Science (CLOSER 2017), pages 368-375

ISBN: 978-989-758-243-1

ployed as a virtual machine or container

restricted

to operate onto a subset of the cores actually available

on the physical server, with a reduced overall amount

of RAM, local persistent storage, and even restricted

networking capabilities (throttling). However, while

memory partitioning and network throttling can be

done with quite a degree of ﬂexibility, partitioning of

physical cores constitutes a major restriction on the

possible available choices, and the number of func-

tions that can be deployed onto the same system. In

many cases, to enhance resources utilization, physi-

cal cores are shared among multiple virtualized func-

tions, but there are no ways to control exactly how

much time the shared processors are dedicated to each

such function, resulting in services suffering of very

variable and unpredictable performance ﬂuctuations.

Contributions. In this paper, the problem of tem-

poral isolation among a number of network functions

co-located on the same physical node is tackled. A

design is proposed for a multi-core server capable

of hosting multiple functions with predictable perfor-

mance levels, leveraging two main building blocks:

a) use of a reservation-based real-time scheduler at

the OS kernel level, providing scheduling guarantees

to, and temporal isolation across, individual func-

tions; b) an insightful queueing-based mathematical

model that links performance statistics that can be ex-

pected from the system with its conﬁguration options

in terms of scheduling parameters and input trafﬁc

characterization. The CPU scheduler in use allows for

guaranteeing each deployed function both the com-

putational bandwidth and the precise time granular-

ity over which CPU cycles are granted. The model

is veriﬁed through experimental validation made on a

synthetic workload and scenario inspired to real data

coming from an LTE use-case.

2 PROBLEM PRESENTATION

2.1 NFV and Virtualized RAN

In the context of radio access networks (RANs), the

NFV paradigm is translating into a number of solu-

tions based on the concept of virtualizing RAN func-

tions (Clo, 2015; Haberland et al., 2013; Costa-Perez

In what follows, we refer to the term virtualized func-

tion and virtual machine in a loose sense, including the pos-

sibility that these functions are deployed as virtual machines

on a hypervisor, or simply deployed within different con-

tainers or isolated processes within the same operating sys-

tem or runtime environment.

L2 low

REC

L2 high

Figure 1: Possible RAN/VRAN deployment models.

et al., 2013; Maaref et al., 2014; Cis, 2016). In these

architectures, packet processing functions are moved

away from the antenna equipment and consolidated

into geo-distributed private-cloud data centers. For

example, in the context of LTE (see Figure 1), there

have been proposals to move both the physical (L1)

and the MAC (L2) layers to a shared data center, let-

ting antennas become merely remote radio heads con-

nected to a centralized Remote Equipment Controller

(REC) in the private cloud via high-performance con-

nectivity such as CPRI (CPR, 2015). However, such

an approach poses very high bandwidth requirements

and stringent-to-prohibitive round-trip latency con-

straints (in the order of 100µs) on the CPRI link. A

certainly more viable approach is the one to central-

ize in the NFV infrastructure only the MAC layer

processing, where the requirement on round-trip la-

tency becomes way more satisﬁable, e.g., air inter-

face L2 packets need to be processed in less than

6ms (Cis, 2016), where 3.66ms is a typically accept-

able L2 round-trip latency (Pichon, 2014).

Also, L2 processing functions are usually CPU-

intensive, e.g., as they usually involve encryp-

tion/decryption of data. Therefore, traditional ap-

proaches to fulﬁlling the tight latency/processing con-

straints resort to dedicated hardware accelerators in

such cases. However, novel cloud-based paradigms

usually rely on commodity hardware so it makes

sense to focus on software-only solutions.

2.2 Problem Statement

Consider the scenario in Figure 2: several network

functions are deployed onto a single Digital Unit, per-

forming L2 packet processing functions for LTE cel-

lular networks. We focus on controlling the temporal

interference as due to co-scheduling of CPU-intensive

network functions on the same cores and modeling

their resulting performance. Therefore, interferences

due to competing access to other resources than the

CPU, like network or disk or cache/memory hierar-

Temporal Isolation Among LTE/5G Network Functions by Real-time Scheduling

369

chy, is post-poned to future extensions of the work.

Trafﬁc coming from each cell is handled by one

or multiple threads that need to process each received

packet within a response-time R

∗

, at least for a (high)

percentage of the trafﬁc φ. In other words, denot-

ing with R the stochastic variable representing the

response-time of each of our L2 trafﬁc processing

functions, the probabilistic Service-Level Objective

(SLO) under consideration is formally stated as:

{

R ≤ R

∗

}

≥ φ. (1)

The considered reference system is composed of

a single server with m identical cores, where multi-

ple worker threads are deployed, serving trafﬁc from

a number of cells. The considered scenario and anal-

ysis shown in Section 4 deals explicitly with the case

of n workers serving trafﬁc from the same cell, with

a random load-balancing of the trafﬁc among them.

Input trafﬁc from each cell is assumed to be a Pois-

son process, with independent identically and expo-

nentially distributed (i.i.d.) inter-arrival times, with

average aggregate arrival rate λ

and average inter-

arrival times of

. Processing times for workers are

assumed exponentially distributed and i.i.d. as well,

with average service rate µ and average computation

time

, when processed in isolation on a single core.

Note that it is assumed that the platform does not have

CPU frequency switching capabilities, or that they are

disabled if present, otherwise the input trafﬁc volume

would inﬂuence the packet processing times. Indeed,

in the majority of modern architectures, static power

consumption is dominant w.r.t. dynamic one. Micro-

processors based on such technologies save consider-

able more energy when in deep-sleep states, so it is

convenient to run them at the maximum frequency so

that the amount of time they spend in the sleep state

is maximized (Benini et al., 2000).

Traffic

Generator

Figure 2: Reference scenario: a cell trafﬁc is processed by

one or more workers deployed on a multi-core server.

In this paper, a model is shown that, given a char-

acterization of the input process, allows for tuning the

scheduling conﬁguration for the workers so as to meet

the SLO requirements of Equation (1). As it will be

clear shortly, the presented closed-form model pro-

vides also additional insights to system designers.

3 RESERVATION-BASED CPU

SCHEDULING

Resource reservation scheduling (Mercer et al., 1993)

is a well-known model for sharing a physical

CPU/core among a number of tasks, in a way that

each one of them can be granted a pre-ﬁxed amount

of time Q on the CPU (called budget) every allocation

window of duration P (called scheduling period), pro-

viding temporal isolation among co-located tasks. A

speciﬁc algorithm implementing this paradigm is the

Constant Bandwidth Server (CBS) (Abeni and But-

tazzo, 1998), a technique that gained attention from

the Linux community recently as due to its availabil-

ity within the mainline kernel since version 3.14 (Lelli

et al., 2016), in the form of the SCHED DEADLINE

scheduling policy.

Under speciﬁc conditions (see

Section 3), this allocation of the CPU is granted

by the scheduler, independently of how other tasks

behave in the system, effectively providing tempo-

ral isolation among them. On a multi-core systems,

SCHED DEADLINE can provide both global and parti-

tioned EDF (G-EDF and P-EDF, respectively).

SCHED DEADLINE adds to the real-time schedul-

ing features of the Linux kernel in serving soft real-

time workloads, making it an increasingly appeal-

ing platform for the execution of virtualized network

functions with predictable performance levels (Vi-

tucci et al., 2014; Clo, 2015).

Schedulability Guarantees. SCHED DEADLINE is

known to provide temporal isolation among worker

tasks with no deadline misses if either of the follow-

ing conditions is satisﬁed: 1) if tasks are globally

scheduled, reservations respect the G-EDF schedula-

bility bound (Goossens et al., 2003) for m cores,

∑

i=1

≤ m(1 −U

max

) +U

max

, (2)

where U

max

, max

i=1,..., n

; if Q

= U ∀i:

U ≤

n − 1 + m

; (3)

if n <= m G-EDF will migrate each task to a differ-

ent core with the only constraint that Q

≤ 1 for

each i; 2) if tasks are partitioned among cores, reser-

vations need to respect the P-EDF (single-processor)

schedulability bound (Liu and Layland, 1973): as-

suming each core j hosts n

workers with

∑

j=1

= n:

∑

i=1

j, i

≤ 1, for each j = 1, . . . , m. (4)

https://www.kernel.org/doc/Documentation/

scheduler/sched-deadline.txt

CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science

370

The traditional way to enforce timing constraints

and full isolation assigns

{

}

equal to the task worst-

case computing time and

{

}

equal to the minimum

inter-arrival times for subsequent requests. However,

such an approach is aimed at hard real-time systems,

where even a single deadline miss cannot be tolerated.

In this paper, a soft real-time approach is under-

taken instead, where the system has to satisfy precise

timing requirements expressed as a probabilistic SLO.

Therefore, it is possible to trade exact schedulability

for efﬁciency, operating the system at a point that is

beyond the above teheoretical schedulability bound,

especially for the G-EDF case, because the test in

Equation (2) is quite pessimistic (see Section 5).

4 APPROACH AND ANALYSIS

In this section, a probabilistic response-time analy-

sis based on queueing theory is presented, for the

LTE/5G application scenario described in Section 2.

4.1 SCHED DEADLINE Queueing Model

In the proposed architecture, network func-

tions are deployed by scheduling them using

SCHED DEADLINE, with scheduling parameters

{

, P

)

}

i=1

to be decided as detailed below. Thanks

to the temporal isolation property enforced by

SCHED DEADLINE (see Section 3), each worker

task i with a SCHED DEADLINE runtime of Q

and

period of P

≥ Q

can be seen as an isolated queueing

system whose service rate µ

has a slow-down factor

w.r.t. the maximum sequential service rate µ:

µ (5)

If Poissonian trafﬁc from a cell is randomly spread

across more workers so that a percentage p

of the

aggregate trafﬁc λ

is picked up by worker i, then

sub-ﬂows in input to the workers are still Poissonian

with an average rate of λ

= p

· λ

. Since both ar-

rival and service rates are assumed to be exponen-

tially distributed, the behavior of each worker can be

approximated as an isolated M/M/1 queue (Kendall,

1953). Refer to Section 5 for an experimental vali-

dation of the approximation and the derived results.

General M /M/1 results (Gross et al., 2008) are spe-

cialized to our case in what follows, where for sim-

plicity the discussion is referred to a single cell whose

entire aggregate trafﬁc λ

is split across the n workers

to be deployed onto an m-cores platform.

Let ρ be deﬁned as ρ ,

≡

nQµ

and assumed

constant across all workers (to ease notation, the i in-

dex is omitted, when the discussion is implicitly refer-

ring to a speciﬁc worker or all workers have the same

conﬁguration). First, let us consider the schedulabil-

ity condition for the scheduler as detailed in Section 3.

Under P-EDF, assuming a system ﬁlled up with re-

served tasks, the maximum theoretical computational

bandwidth is derived from Equation (4):

=⇒ ρ =

. (6)

Under G-EDF, instead, Equation (2) leads to:

n − 1 + m

=⇒ ρ =

n − 1 + m

. (7)

In addition to the theoretical schedulability condi-

tions, it is useful to make explicit the general stability

condition for the system, e.g., that incoming trafﬁc on

each queueing system on average does not exceed the

server processing capability:

ρ < 1 =⇒

nµ

. (8)

Once the servers are properly isolated, each of

them behaves as an independent M/M/1 queue, thus

the average response time R

and waiting time W

are

obtained as:

E [R

] = E [W

] +

− λ

. (9)

The response-time R

distribution is known to be ex-

ponential as well with parameter µ

− λ

, thus we get:

• R

probability distribution function (PDF):

(t) = (µ

− λ

−(µ

−λ

for each t > 0, (10)

• R

cumulative distribution function (CDF):

(t) = 1 − e

−(µ

−λ

for each t > 0. (11)

4.2 Choice of the (Q

) Parameters

The formal SLO requirement for the system, as given

in Equation (1), can now be described as: “φ

per-

centile of the response-time R ≤ a given threshold

∗

”. This requirement can be formalized as follows:

∗

) ≥ φ, (12)

where F

is the response-time CDF introduced in Sec-

tion 4. In the case of an M/M/1 system with param-

eters µ

and λ

, Equation (12) becomes:

≥



−

ln (1 − φ)

∗



. (13)

Note that any (Q

, P

) pair satisfying the proba-

bilistic SLO in Equation (13), satisﬁes also the M/M/1

Temporal Isolation Among LTE/5G Network Functions by Real-time Scheduling

371

stability condition in Equation (8). On the other hand,

schedulability of the reservations on the server has to

be ensured separately, e.g., according to Equations (7)

or (6) if G-EDF or P-EDF is used, respectively.

From Equation (13), the full speciﬁcation of Q

and P is obtained observing that it is convenient to set

P = R

∗

, to avoid temporal interferences and not cause

excessive scheduling on the platform. The point is

further detailed in Section 5.2.

4.3 Maximum Sustainable Arrival Rate

The M/M/1 model described in Section 4 allows us

to compute the maximum sustainable aggregate ar-

rival rate for a given platform conﬁguration: from

Equation (12), we have

= nλ

≤ n



µ +

ln(1 − φ)

∗



, (14)

where the logarithm brings a negative contribution.

We remark that the obtained upper-bound to λ

permits achieving the target response-time percentile,

but it does not guarantee the full temporal isolation

of the reservation servers. Temporal isolation can

be guaranteed by considering the maximum com-

putational bandwidth

saturating the system under

P-EDF in Equation (6), that is:

≤ mµ +

∗

ln(1 − φ) . (15)

On the other hand, using the maximum computational

bandwidth as from theoretical G-EDF schedulability

in Equation (7), we obtain:

≤

n − 1 + m

µ +

∗

ln(1 − φ) . (16)

5 EXPERIMENTAL VALIDATION

In this section, experimental results are presented that

validate the analysis in Section 4. All experiments

have been conducted on a Freescale LS2085A RDB

(ARMv8) board, with 8 cores running at 1.8GHz

and 16GB of RAM, a promising platform for energy-

efﬁcient data-plane processing and NFV. The system

boots an image with Linux 4.1.8, built using Yocto

A synthetic application was built around the LTE sce-

nario described in Section 2: a client acting as traf-

ﬁc generator is capable of submitting requests accord-

ing to tunable exponential inter-arrival times; a multi-

threaded server is able to serve requests received from

the client, taking a tunable exponentially distributed

https://www.yoctoproject.org/

computation time; these are designed to dump on disk

the full set of observed response-times throughout

each experiment. The server can conﬁgure its threads

using tunable SCHED DEADLINE settings, including

the use of either G-EDF or P-EDF.

5.1 Scenario Parameters

In the considered scenario, n = 8 worker threads have

been instantiated, scheduled under G-EDF across m =

4 out of the available cores, and conﬁgured to serve

trafﬁc with a packet processing rate of µ

∼

5300pkt/s.

A single client generator submitted requests at an av-

erage rate of λ

= 15000pkt/s, obtaining an indi-

vidual (worker) trafﬁc average rate of: λ

1875pkt/s. The cell trafﬁc type and rate, as well as

the workers service rate, have been chosen by best-

ﬁtting parameters over data gathered from a real LTE

installation. Also, the same industrial use-case man-

dated the needed probabilistic SLO parameters: 99%

of the packets served within 2ms, namely: R

∗

= 2ms

and φ = 0.99. The presented results have been ob-

tained via 1-minute long runs, each relying on 15000∗

60 = 900K measured response times.

5.2 Experimental Results

Theoretical results from Section 4, speciﬁcally Equa-

tions (13) and (8), can easily be applied for the just

explicited conﬁguration, obtaining, among others:

• min Q/P for stability

nµ

∼

35.38%;

• min Q/P for SLO:

≥

−

ln(1−φ)

∗

∼

78.8%.

These results have been compared with the obtained

experimental data, shown in Figure 3, where the ob-

tained response-time 99

percentile (on the Y axis)

is plotted with various Q/P ratios (on the X axis),

and various scheduling periods (different data sets).

The probabilistic SLO requirement is clearly satisﬁed

when

≥ 0.7, as per the theoretical expectation.

2000

4000

6000

8000

10000

12000

14000

0.4 0.5 0.6 0.7 0.8 0.9

Time (us)

Q/P

Experimental response-time p99

P: 250 us

P: 500 us

P: 1000 us

P: 2000 us

Figure 3: Experimental thresholds (R

∗

) when φ = 0.99, for

different values of the ratio Q

CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science

372

Selection of the Period P

. Equation (13) provides

a choice for the Q/P ratio, for a worker that needs to

service trafﬁc with precise probabilistic guarantees.

The exact parameters can be determined based on

the following considerations, focusing on the choice

of the period P. First, a smaller period leads to in-

creased scheduling overheads and context switches

(roughly 2 per scheduling period per task); indeed,

the obtained experimental 90

and 99

percentiles of

the response times were slightly better (lower) with

higher periods, among the ones tried (see Figure 3).

Second, the scheduling period determines the time

granularity based on which SCHED DEADLINE sched-

ules each reserved task, thus the scheduling delay is

roughly bounded by P; indeed, experimental results

show slightly worse worst-case response-times when

using higher periods. As also found experimentally,

setting P = 2ms allows for meeting the temporal re-

quirements without excessive overheads.

In another 1-min experiment, the context switches

in the platform throughout the whole run have been

counted, in order to highlight how many additional

context switches can happen with too small of a pe-

riod. Results are reported in Figure 4. With a pe-

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

0 500 1000 1500 2000

Context switches (x10

)

Period when using SCHED_DL

SCHED_DL

SCHED_FIFO

Figure 4: Number of context switch for SCHED DEADLINE

and SCHED FIFO, when varying the reservation periods.

riod of 2ms, SCHED DEADLINE achieves a reduc-

tion in the number of context switches of about 10%

compared to SCHED FIFO. However, using arbi-

trarily small periods can lead SCHED DEADLINE to

overwhelm the platform with an excess of context

switches (e.g., +40% extra context switches with a

100µs period).

Temporal Isolation. The advantage of using

SCHED DEADLINE resides in its capability to provide

temporal isolation with a speciﬁc time granularity

among workers. In this section, this is highlighted

through an experiment comparing SCHED

DEADLINE

vs SCHED FIFO in presence of an anomalous trafﬁc

overloading one of the workers. The overload has

been simulated by using a higher probability of 0.475

for one of the workers and a lower probability of

0.075 for the remaining 7 workers. Then, raising the

overall trafﬁc rate to λ

= 25000pkt/s, we obtained 7

individual rates at the usual 1875pkt/s while the 8th

individual rate is at the higher value of 11875pkt/s.

Budgets and periods for SCHED DEADLINE have

been set to 900µs and 1000µs, respectively.

100

0 500 1000 1500 2000

Probability (0-100)

Response time (us)

SCHED_DL

SCHED_FIFO

Figure 5: Response time CDFs in the presence of overload,

obtained under the scenario of Section 5.2.

Figure 5 compares the response-time CDF ob-

tained for one of the regular workers when us-

ing SCHED DEADLINE and when using SCHED FIFO.

The response times are higher under SCHED FIFO,

due to the interference of the overloaded worker, ex-

hibiting a +300us increase in the 99th percentile of

the response times compared to the ones obtained un-

der SCHED DEADLINE. On the other hand, the over-

loaded worker managed to serve 26% more packets

under SCHED FIFO, because its computing capabili-

ties were not restricted by the scheduling budget as in

the case of SCHED DEADLINE, as expected.

6 RELATED WORK

In the context of the Long Term Evolution (LTE) tech-

nology, virtualization-based mobile network architec-

tures are now widely accepted, and NFV is now rec-

ognized as an established solution to be deployed in

core networks and data centers (Zaki et al., 2011;

Sama et al., 2015; Nguyen et al., 2016). However,

scarce attention has been dedicated so far to the prob-

lem of QoS assurance in edge nodes. This paper tries

to ﬁll this gap by addressing predictable execution

within edge nodes to ensure QoS control and tem-

poral isolation, by resorting to an EDF-based real-

time scheduling strategy. On a related note, in (Jansen

and Mullender, 2003) authors propose to apply EDF-

based real-time scheduling within a Plan9-based ex-

perimental base-station software stack, albeit the fo-

cus of the work is on preventing priority inversion

through a speciﬁc deadline inheritance mechanism.

Temporal Isolation Among LTE/5G Network Functions by Real-time Scheduling

373

A similar approach has been more recently presented

in (Vitucci et al., 2014), but no attempt has been made

to build a precise model of the function that takes

properly into consideration the scheduling parame-

ters, as done in this paper.

The problem of resource partitioning for tempo-

ral isolation has been largely investigated in the real-

time scheduling literature. The main concern in this

regard was to ensure that multiple software compo-

nents could be deployed on the same physical cores

ensuring that any task overrun could not affect the

execution of other tasks and disrupt system integrity.

This need mainly arises whenever a timely service has

to be ensured in a system with heterogeneous timing

requirements or dynamic computational workload.

To achieve this property, several different Resource

Reservation algorithms have been developed. The

earliest examples (Rajkumar et al., 1998; Abeni and

Buttazzo, 1998) have been developed in the context

of multimedia applications, with the intent to offer

predictable levels of quality-of-service (QoS) to real-

time workload in highly heterogeneous environments.

Resource reservation techniques, which had origi-

nally been proposed for the execution of indepen-

dent tasks on single-core systems, were also extended

to cope with hierarchical scheduling systems (Feng

and Mok, 2002) and multi-core platforms (Shin et al.,

2008; Easwaran and Andersson, 2009).

The effectiveness of these scheduling mechanisms

is also supported by the recent introduction of the

SCHED DEADLINE scheduling class in the Linux ker-

nel (Lelli et al., 2016). Traditional Linux schedul-

ing classes either do not provide real-time guaran-

tees (see SCHED OTHER policy, which schedules ac-

tivities in a best-effort way), or implement priority-

based real-time scheduling with no resource reserva-

tion capabilities (see SCHED FIFO or SCHED RR poli-

cies). Vice versa, SCHED DEADLINE allows com-

bining traditional multi-core scheduling with resource

reservations using a variant of the Constant Band-

width Server (CBS) algorithm (Abeni and Buttazzo,

1998) to obtain temporal isolation among tasks. In-

terestingly, SCHED

DEADLINE does not require any

particular assumption on tasks characteristics, allow-

ing it to serve highly heterogeneous workloads.

A remarkable work in this area has been carried

out in the context of the IRMOS European project

where an implementation of the EDF/CBS algorithm

within the Linux kernel has been used to provide

strong scheduling guarantees to KVM-based VMs

running in a virtualized real-time cloud infrastructure.

However, one of the critical issues that emerged was

the one of tuning proper values for the scheduling pa-

http://http://www.irmosproject.eu/

rameters, and how to link them to the overall perfor-

mance of the deployed application. In this work, we

mark a signiﬁcant step forward in such an issue, de-

veloping a closed-form mathematical model capable

of capturing the relationship between system conﬁgu-

ration, input trafﬁc characterization and expected per-

formance of the virtualized network functions.

7 CONCLUSIONS AND FUTURE

WORK

In this paper, an approach has been proposed to

provide temporal guarantees to LTE trafﬁc pro-

cessing functions, based on the SCHED DEADLINE

Linux scheduler. This allows for modelling the per-

formance of individual functions applying classical

queueing theory. Experimental data shows that the

proposed model matches the application behavior.

SCHED DEADLINE allows for controlling the compu-

tational bandwidth ratio dedicated to each function, as

well as the time granularity of the schedule.

The presented model is shown to provide easy-to-

use closed-form formulas under Poissonian trafﬁc as-

sumptions, providing insightful relationships among

parameters related to the system and scheduler con-

ﬁguration, and expected performance ﬁgures.

The presented work has various limitations that

need to be addressed in additional future research.

First, the proposed methodology needs to be applied

in a more realistic scenario, where multiple cells sub-

mit trafﬁc to functions deployed throughout a dis-

tributed NFV infrastructure. Second, the paper fo-

cused on CPU-dominated workloads, but in real life,

there will also be a mix of data-intensive activities that

saturate disk access or available NIC bandwidth, be-

fore the CPU. The modeling and analysis technique

needs to be extended to cope with these more involved

scenarios. Furthermore, it would be interesting to ver-

ify what integration issues arise for integrating the

reservation scheduler paradigm within a real NFV in-

frastructure. This might be done, e.g., extending OP-

NFV

or other frameworks.

REFERENCES

(2015). Cloud RAN – Ericsson White Paper.

www.ericsson.com/res/docs/whitepapers/wp-

cloud-ran.pdf.

(2015). Common Public Radio Interface (CPRI) v7.0 –

https://www.opnfv.org/

CLOSER 2017 - 7th International Conference on Cloud Computing and Services Science

374

Interface Speciﬁcation. http://www.cpri.info/

downloads/CPRI_v_7_0_2015-10-09.pdf.

(2016). Cisco 5G Vision Series: Small Cell Evolution.

Cisco Whitepaper.

Abeni, L. and Buttazzo, G. (1998). Integrating multimedia

applications in hard real-time systems. In Proc. IEEE

Real-Time Systems Symposium, pages 4–13, Madrid,

Spain.

Armbrust, M., Fox, A., Grifﬁth, R., Joseph, A. D., Katz,

R. H., Konwinski, A., Lee, G., Patterson, D. A.,

Rabkin, A., Stoica, I., and Zaharia, M. (2009). Above

the clouds: A berkeley view of cloud computing.

Technical Report UCB/EECS-2009-28, EECS De-

partment, University of California, Berkeley.

Benini, L., Bogliolo, A., and Micheli, G. D. (2000). A

survey of design techniques for system-level dynamic

power management. IEEE Transactions on Very Large

Scale Integration (VLSI) Systems, 8(3):299–316.

Costa-Perez, X., Swetina, J., Guo, T., Mahindra, R., and

Rangarajan, S. (2013). Radio access network virtual-

ization for future mobile carrier networks. IEEE Com-

munications Magazine, 51(7):27–35.

Easwaran, A. and Andersson, B. (December 2009). Re-

source sharing in global ﬁxed-priority preemptive

multiprocessor scheduling. In Proceedings of the 30th

IEEE Real-Time Systems Symposium (RTSS), Wash-

ington, DC, USA.

Feng, X. and Mok, A. K. (December 2002). A model of

hierarchical real-time virtual resources. In Proceed-

ings of the 23rd IEEE Real-Time Systems Symposium

(RTSS), Austin, TX, USA.

Fiorani, M., Skubic, B., M

artensson, J., Valcarenghi, L.,

Castoldi, P., Wosinska, L., and Monti, P. (2015). On

the design of 5G transport networks. Photonic Net-

work Communications, 30(3):403–415.

Goossens, J., Funk, S., and Baruah, S. (2003). Priority-

driven scheduling of periodic task systems on multi-

processors. Real-Time Systems, 25(2):187–205.

Gross, D., Shortle, J. F., Thompson, J. M., and Harris, C. M.

(2008). Fundamentals of Queueing Theory. Wiley-

Interscience, 4th edition.

Haberland, B., Derakhshan, F., Grob-Lipski, H., Klotsche,

R., Rehm, W., Schefczik, P., and Soellner, M. (2013).

Radio base stations in the cloud. Bell Labs Technical

Journal, 18(1):129–152.

Hawilo, H., Shami, A., Mirahmadi, M., and Asal, R. (2014).

NFV: state of the art, challenges, and implementation

in next generation mobile networks (vepc). IEEE Net-

work, 28(6):18–26.

Jansen, P. G. and Mullender, S. (2003). Real-time in plan 9

: a short overview.

Kendall, D. G. (1953). Stochastic processes occurring in

the theory of queues and their analysis by the method

of the imbedded markov chain. The Annals of Mathe-

matical Statistics, 24(3):338–354.

Lelli, J., Scordino, C., Abeni, L., and Faggioli, D. (2016).

Deadline scheduling in the linux kernel. Software:

Practice and Experience, 46(6):821–839.

Liu, C. L. and Layland, J. W. (1973). Scheduling algo-

rithms for multiprogramming in a hard-real-time en-

vironment. Journel of the ACM, 20(1):46–61.

opez-P

erez, D., Ding, M., Claussen, H., and Jafari,

A. H. (2015). Towards 1 gbps/ue in cellular sys-

tems: Understanding ultra-dense small cell deploy-

ments. IEEE Communications Surveys Tutorials,

17(4):2078–2101.

Maaref, A., Ma, J., Salem, M., Baligh, H., and Zarin, K.

(2014). Device-centric radio access virtualization for

5g networks. In 2014 IEEE Globecom Workshops (GC

Wkshps), pages 887–893.

Mercer, C. W., Savage, S., and Tokuda, H. (1993). Proces-

sor Capacity Reserves: An Abstraction for Managing

Processor Usage. In Proc. 4th Workshop on Worksta-

tion Operating Systems.

NFV Industry Specif. Group (2012). Network Functions

Virtualisation. Introductory White Paper.

Nguyen, V.-G., Do, T.-X., and Kim, Y. (2016). Sdn and

virtualization-based lte mobile network architectures:

A comprehensive survey. Wireless Personal Commu-

nications, 86(3):1401–1438.

Pichon, D. (2014). Mobile Cloud Networking FP7 Euro-

pean Project: Radio Access Network as a Service. 4th

Workshop on Mobile Cloud Networking. Available

at: http://docplayer.net/3714038-Mobile-

cloud-networking-fp7-european-project-

radio-access-network-as-a-service.html.

Rajkumar, R., Juvva, K., Molano, A., and Oikawa, S.

(January 1998). Resource kernels: A resource-

centric approach to real-time and multimedia systems.

In SPIE/ACM Conference on Multimedia Computing

and Networking, San Jose, CA, USA.

Sama, M. R., Contreras, L. M., Kaippallimalil, J., Akiyoshi,

I., Qian, H., and Ni, H. (2015). Software-deﬁned con-

trol of the virtualized mobile packet core. IEEE Com-

munications Magazine, 53(2):107–115.

Shin, I., Easwaran, A., and Lee, I. (July 2008). Hierarchical

scheduling framework for virtual clustering of multi-

processors. In 20th Euromicro Conference on Real-

Time Systems (ECRTS), Prague, Czech Republic.

Vitucci, C., Lelli, J., Parri, A., and Marinoni, M. (2014). A

Linux-based Virtualized Solution Providing Comput-

ing Quality of Service to SDN-NFV Telecommunica-

tion Applications. In Proceedings of the 16th Real

Time Linux Workshop (RTLWS 2014), pages 12–13,

Dusseldorf, Germany.

Zaki, Y., Weerawardane, T., Gorg, C., and Timm-Giel, A.

(May 2011). Multi-qos-aware fair scheduling for LTE.

In 73rd IEEE Vehicular technology conference (VTC-

Spring), Budapest, Hungary.

Temporal Isolation Among LTE/5G Network Functions by Real-time Scheduling

375