The Importance of Being OS-aware

∗

In Performance Aspects of Cloud Computing Research

Tommaso Cucinotta

, Luca Abeni

, Mauro Marinoni

and Carlo Vitucci

Scuola Superiore Sant’Anna, Pisa, Italy

Ericsson, Stockholm, Sweden

Keywords:

Cloud Computing, Operating Systems, Quality of Service Control, Temporal Isolation, Real-Time Scheduling.

Abstract:

This paper highlights inefﬁﬁencies in modern cloud infrastructures due to a distance between the research

on high-level cloud management / orchestration and the research on low-level kernel and hypervisor mecha-

nisms. Our position about this issue is that more research is needed to make these two worlds talk to each

other, providing richer abstractions to describe the low-level mechanisms and automatically map higher-level

descriptions and abstractions to conﬁguration and performance tuning options available within operating sys-

tems and kernels (both host and guest), as well as hypervisors.

1 INTRODUCTION

Despite the tremendous growth and transformation

undergone by Cloud Computing over the last 10

years, this computing paradigm ﬁnds its roots back

in the ’60s. The paradigm of utility computing was

mentioned in the MIT centennial speech by John Mc-

Carthy in 1961 (Garﬁnkel, 2011), in its famous state-

ment: “Computing may someday be organized as a

public utility just as the telephone system is a pub-

lic utility”. Also, machine virtualization, one of the

core enablers of cloud computing, had been concep-

tualized and prototyped already in the ’70s (White,

2015), albeit the hardware available at that time was

not ready for a widespread use of the mechanism.

Furthermore, since the Internet was born, we have

seen web-hosting servers on rental, sometimes along

with simple accompanying data-base and back-up ser-

vices. However, the actual possibility for the general

public to be able to rent virtual machines (VMs) on-

demand, rapidly and in a self-service and completely

automated fashion, dates back to year 2006, when

Amazon Web Services (AWS) launched its Elastic

Compute (EC2) service (Information Resources Man-

agement Association, 2012).

At that time, you could rent simple virtual ma-

chines choosing among a few instance types (Barr,

2015), using an Infrastructure-as-a-Service model.

∗

This work was partially funded by Ericsson AB, Stock-

holm, Sweden.

Later, other players entered the market, and the

available services evolved up to the nowadays na-

tive cloud applications making use of a plethora

of services available from a Platform-as-a-Service

(PaaS) provider, like security, load-balancing, high-

availability, data storage and replication, distributed

communication middleware and queueing, big-data

processing frameworks, etc. Recently, operating-

system (OS) level virtualization, a.k.a., contain-

ers, are increasingly replacing machine virtualiza-

tion in various application domains, ﬁnding a natural

ﬁt within novel distributed processing architectures

based on microservices (Brown, 2016).

Academic research has been ﬂanking the emerg-

ing industry of cloud computing services, experiment-

ing with novel directions of investigation, mostly in-

spired to the consolidated areas around distributed

computing, fault-tolerance, data-bases, networking

protocols, routing, data replication and distributed

transactions, security, etc., with many of these re-

search areas overﬂowing almost naturally from prior,

consolidated research lines in the ﬁeld of grid and

distributed computing. Many research groups fo-

cused on novel issues coming straight from particu-

lar characteristics of the cloud computing paradigm

instead, such as the capability to grow and shrink ser-

vice instantiations by adding or removing VMs at run-

time, live-migrate them for re-optimization purposes,

as well as the one to create virtual networking over-

lays on top of the physical infrastructure. This re-

sulted in novel research on scalable resource manage-

626

Cucinotta, T., Abeni, L., Marinoni, M. and Vitucci, C.

The Importance of Being OS-aware.

DOI: 10.5220/0006793906260633

In Proceedings of the 8th International Conference on Cloud Computing and Services Science (CLOSER 2018), pages 626-633

ISBN: 978-989-758-295-0

ment in massively distributed infrastructures, dealing

with issues related to monitoring, resource estimation

and workload prediction (Cortez et al., 2017), optimal

VM placement (Chang et al., 2010; Mann, 2015) in-

cluding cases of network-aware placement (Alicherry

and Lakshman, 2012; Yao et al., 2013; da Silva and

da Fonseca, 2016; Steiner et al., 2012) and allocation

of data-intensive VMs (Zhang et al., 2015), energy

efﬁciency (Ghribi et al., 2013), elastic applications

management and associated performance models (Xia

et al., 2015), architectures of cloud orchestration lay-

ers, and others.

Just recently, further ﬁelds of particular success in

terms of research productivity have been the ones of

cloud federation (Konstanteli et al., 2014) and edge

(cloud) computing. The last one, in particular, in-

cludes cloudlets (Satyanarayanan et al., 2009) and

the cloudiﬁcation of other IT areas such as those

of interest for network operators, with the uprising

trend of software deﬁned networking (SDN) and net-

work function virtualization (NFV) (Public, 2016;

Networks, 2017) and related technologies (Fiorani

et al., 2015). These application scenarios are charac-

terized by tighter constraints concerning low latency

and overall Quality of Service (QoS) for mobile ap-

plications (Valcarenghi et al., 2017). Moreover, the

development of end-to-end SDN-NFV solutions can-

not ignore the characteristics and, therefore, the needs

of Virtual Functions (VF) that have to be placed in

nodes located in different zones of the network, such

as the core or the edge. VFs assigned to core nodes

present similar needs to the classic cloud infrastruc-

tures, while the closer you get to the edge of the

network, the more the demand for ﬂexible manage-

ment of the computing resource becomes critical (Vi-

tucci and Larsson, 2017). These requirements are cru-

cial both concerning the QoS levels and the available

resources optimization because, in such contingen-

cies of scarce resources and signiﬁcant power con-

sumption (SCTE, 2016), a more tailored allocation of

energy and resources becomes essential (Gai et al.,

2016).

2 THE ROLE OF THE

OPERATING SYSTEM

As due to the intrinsic nature of cloud computing,

performance of cloud-hosted applications, services,

and internal infrastructure management systems has

always been of paramount importance, in the overall

picture. Similarly to when one buys physical hard-

ware, renting on-line IaaS facilities needs to respect

minimum performance speciﬁcations that are usually

made more or less explicit by the provider. However,

plenty of the mentioned research works tackling per-

formance and QoS control aspects of cloud systems,

have been focusing on exploiting elasticity to just

compensate for the great variability in performance

of the rented infrastructure items (unstable VMs pro-

cessing performance, usually, due to time-sharing of

the physical CPUs, but also instability in disk ac-

cess as well as networking performance). There-

fore, one ﬁnds an increasing presence of high-level re-

source management frameworks, orchestration layers

and evolved middleware services, dealing with work-

load monitoring, estimation and prediction, as well as

resource management.

One aspect that has not been given sufﬁcient at-

tention, in our opinion, is the bottommost part of the

software stack: the operating system (OS) and hyper-

visor layers. Indeed, a signiﬁcant amount of research

in the ﬁeld focused on the higher levels of the cloud

stack, just relying on well-established basic function-

alities provided by traditional OSes, ﬁrst and fore-

most, the ability to pin down virtual cores onto phys-

ical cores of the underlying hardware platform, and

in some cases ﬁne-tuning of the hypervisor memory

allocation subsystem by exploiting the NUMA capa-

bilities of the hardware. However, OSes and hypervi-

sors have been evolving over time, exhibiting richer

and richer functionality for supporting ﬁne-tuning of

performance, with reference to achievable latency and

throughput within data centre infrastructures, encom-

passing multiple areas, such as: CPU scheduling,

memory bandwidth allocation and management, in-

terrupt dispatching and virtualized interrupt handling,

para-virtualization options and ways to access hard-

ware accelerators including GPUs, GP-GPUs, crypto-

graphic accelerators or generic FPGA-based process-

ing elements (Caulﬁeld et al., 2016; Pellerin, 2017).

As exempliﬁed in Figure 1, these capabilities are of-

ten pre-tuned according to data centre-wide proce-

dures (or just set as found in OS and other software

distributions), and from the high-level perspective of

infrastructure orchestration layers they are either not

made accessible, or subsumed at a too high of an ab-

straction level, thus resulting in a mismatch with the

plethora of heterogeneous requirements coming from

different applications, resulting in a poor effective-

ness in infrastructure management.

However, it is our belief that the novel panorama

of heavily distributed, elastic, fault-tolerant and of-

ten interactive, computing applications and services

calls for novel OS services and low-level middleware,

and that these need to be better connected to higher-

level cloud management and orchestration layers. Us-

ing more advanced functionalities provided by mod-

The Importance of Being OS-aware

627

Figure 1: Limits of high-level infrastructure orchestration

layers in handling the plethora of low-level performance

tunables available in nowadays OSes and hypervisors.

ern OSes can signiﬁcantly improve the performance,

scalability, predictability thus usability of cloud ser-

vices, making them even more attractive.

3 LOW LEVEL SUPPORT

This section brieﬂy introduces some of the advanced

features that can be provided at the lower-levels of

the cloud stack, and that could be exploited by higher

level cloud control infrastructures to improve perfor-

mance.

3.1 CPU Scheduling

Management of the CPU resources in cloud infras-

tructures follows basically two main patterns: the best

VM instance types are mapped 1-to-1 so that virtual

CPUs run onto dedicated physical CPUs. On the other

hand, the cheapest VM instance types are mapped

onto clusters of physical machines where there is

an over-allocation of virtual CPUs, compared to the

available physical CPUs. Finally, the most expen-

sive and best isolated instance types are those rely-

ing on entire physical machines dedicated to a single

tenant’s instance, either in virtualized or also in bare-

metal form.

Focusing on physical machines shared among

multiple tenants, the dedicated 1-to-1 virtual to physi-

cal core mapping is the set-up that tries to achieve the

best temporal isolation among different VMs. How-

ever, this is not sufﬁcient to isolate data-intensive ap-

plications, such as high-deﬁnition multimedia (video)

processing workloads, big-data analytics and others.

These, even when deployed on dedicated cores, cause

quite a lot of interference to processing tasks deployed

on other physical cores, due to interferences at the

memory access bandwidth and cache-level interfer-

ences. Works aimed at limiting said phenomenon

can be found (Yun et al., 2013), speciﬁcally conﬁn-

ing a precise budget in memory access for individ-

ual physical cores, that, once exhausted, would cause

the VM to be throttled till replenishment of the bud-

get at the next period. Mechanisms have been stud-

ied also to limit and keep under control interferences

at the cache level on big multi-core machines, where

there is a sufﬁciently big last-level cache (LLC). In

such systems, it is both possible to employ software-

only cache partitioning schemes based e.g., on cache

colouring (Kim and Rajkumar, 2016), or to leverage

hardware technologies, such as the recent Intel Cache

Allocation Technology (CAT) (Corbet, 2016).

Whenever virtual CPUs are over-allocated to

physical ones, the way multiple VMs compete among

each other for running on the physical CPUs when ac-

tive at the same time is left to heuristics of the hyper-

visor scheduler, which unfortunately falls short in its

attempt to satisfy requirements of a multitude of ap-

plication domains, without knowing much about what

kind of service/application is running within what

VMs. Normally, the temporal interferences a VM gets

from other co-scheduled VMs are quite high, mak-

ing the performance of each hosted VM quite unsta-

ble. In this area, prior research works focused on the

Xen (Barham et al., 2003) hypervisor, comparing var-

ious scheduler types (Cherkasova et al., 2007) and its

impact on the performance of the hosted workload.

Other research works focused on co-scheduling

virtual CPUs of VMs on the same physical CPUs by

employing a special scheduler within the hypervisor,

that allows for a reservation-based paradigm. Namely

a virtual core may be allotted a budget (a.k.a., run-

time) of Q time units every period of P time units,

so that the underlying hyperivosr (or host OS) sched-

uler guarantees Q time units every period of dura-

tion P, and optionally it also constraints the VM

to be throttled once said budget is exhausted. This

is the example of the IRMOS real-time scheduler

for Linux (Cucinotta et al., 2010; Checconi et al.,

2009), or the similar recent hierarchical extension

of the SCHED DEADLINE scheduler of Linux (Balsini

et al., 2017). Similar efforts have been attempted

on Linux/KVM also by exploiting existing in-kernel

CLOSER 2018 - 8th International Conference on Cloud Computing and Services Science

628

features, so that EDF-based scheduling capabilities

could be injected into the kernel at run-time, without

any need for static code modiﬁcations to the kernel

code (Aesberg et al., 2011).

3.2 Networking

The service performance are affected not only by the

conﬁguration of the used VM, but also by the OS and

kernel mechanisms used to relay data among VMs,

between host and guest applications, etc... For exam-

ple, it has been shown (Abeni et al., 2015) that, when

using Linux as host OS and KVM as hypervisor, sim-

ply changing the packets forwarding mechanism from

the default bridge + tap to macvtap can greatly im-

prove the performance of a virtual router. In partic-

ular, the rate of small packets the virtual router can

forward has been shown to increase of nearly 50%.

Other conﬁguration details (such as the queue length

of the virtual interfaces, the scheduling priority and

interrupt afﬁnity, etc...) that are often ignored in high-

level descriptions also affect the performance and the

ability to sustain a high forwarding rate in a stable

way.

Of course, many of the low-level details are OS

or VM speciﬁc (for example, the possibility to use

macvtap or vhost-net are Linux-only features; Xen

provides a completely different way to forward pack-

ets - with different conﬁguration options - etc....).

This is why many cloud providers often rely to de-

fault settings, that do not require any knowledge of

the OS, kernel, and VM details, resulting, in much

worse performance and usability. As an additional

example, high-performance networking in cloud ser-

vices can often be obtained by using some form of

user-space networking (even arriving to user-space

network drivers). In this regard FreeBSD provides

netmap (Rizzo, 2012a; Rizzo, 2012b), that marked

a signiﬁcant milestone in the area of packet pro-

cessing in general-purpose operating systems, show-

ing how well-engineered optimizations of existing

drivers, coupled with proper novel user-space/kernel-

space interfaces for networking, can lead to a great

increase in processing throughput of small packets.

On different operating systems, similar results can

be obtained by using different mechanisms; for ex-

ample, on Linux-based systems it is possible to use

DPDK

, that has been shown to greatly improve

the performance (especially in software deﬁned net-

works) (Pongrcz et al., 2013). Virtual bridges be-

tween different VMs also can be setup in different

ways, using different technologies that depend on the

used network drivers; for example, if netmap is used

http://www.dpdk.org

then the VALE virtaul switch (Rizzo and Lettieri,

2012) is the best choice for high performance com-

munications among VMs on the same host, reaching

up to 2 million packets per second without hardware

assistance.

Since it is clear that an adminitrator in charge of

setting up a cloud cannot have all of the information

needed to properly select the most appropriate tech-

nologies and conﬁgurations, it becomes clear that it is

of vital importance to automate this process by prop-

erly mapping high-level descriptions understandable

by the administrator to low-level features. To do this,

it is fundamental to develop a way to make the low-

level virtualization infrastructure attributes directly

visible to the higher levels of the cloud management

stack, or exploit richer and more complex abstract in-

terfaces, that allow for properly tuning / conﬁguring

/ using all the available options (even if they are not

standard).

3.3 Computing and Networking

Some authors tried to build mechanisms to improve

the responsiveness of networking applications, in

presence of concurrently running CPU-bound activ-

ities. This is the case of the work in (Kim et al.,

2009), where a heuristic identiﬁes I/O tasks within

a VM, and boosts their scheduling priority (tuning

parameters of the Xen credit-based scheduler) when-

ever there are pending incoming packets to be re-

ceived by that task in the VM. A similar approach has

been attempted on KVM-based systems (Cucinotta

et al., 2011), where a VM being scheduled with a

reservation-based scheduler has been made able to

dynamically switch to a ﬁner-grain scheduling dead-

line whenever there were packets pending to be pro-

cessed by particular services running in the VM at

higher priority.

4 EXPLOITING THE

LOW-LEVEL FEATURES

When the performance of the cloud services becomes

relevant, it is important to properly map high-level

performance requirements of the cloud application or

service to low-level virtualization infrastructure con-

ﬁguration and tuning parameters.

The Importance of Being OS-aware

629

Virtualized

Infrastructure

Manager

VDU

MANO Descriptor

HOT Descriptor

Deployed

VM / Container

reservation (Q,P)

Figure 2: Propagation of the enriched descriptors from the MANO speciﬁcation to the kernel level.

4.1 Adoption of Reservation-based

Approaches

As an example, consider a speciﬁc area of system sup-

port (CPU scheduling): one of the obstacles into get-

ting reservation-based scheduling widely adopted, is

that software design needs a shift from tuning prior-

ities, to tuning budgets and periods, a change that is

not trivial at all. While a reservation period is easily

found considering activation and latency constraints

of an application/service, the budget/runtime is more

difﬁcult to ﬁnd out, also due to its inherent depen-

dency on the speciﬁc platform the software runs atop.

An interesting line of work has been proposed

in the RT-Xen project (Xi et al., 2011), investigat-

ing on the use of various server-based scheduling al-

gorithms for Xen domains, with scheduling parame-

ters set according to sound arguments from hierarchi-

cal real-time theory literature (Feng and Mok, 2002).

RT-Xen, like many of the research works mentioned

above, only focused on low-level kernel or hypervisor

mechanisms, implementing features that do not result

to be available to higher levels of the cloud manage-

ment stack. The only work that tries to address this

concern is RT-OpenStack (Xi et al., 2015), that tries

to integrate the analysis within the OpenStack Nova

scheduling algorithm. While a lot of work still needs

to be done in this area (for example, RT-OpenStack is

strictly dependent on RT-Xen and cannot take advan-

tage of the SCHED DEADLINE (Lelli et al., 2016)

policy that is currently available on vanilla Linux ker-

nels), the RT-Xen / RT-OpenStack work shows a pos-

sible strategy for exploiting advanced features pro-

vided by the lower levels of the cloud stack.

The next subsection shows how a similar approach

can be extended to better integrate with higher level

abstractions and to consider other lower level fea-

tures.

4.2 Bridging Gaps in Abstraction Levels

One of the interesting results shown by RT-OpenStack

is that to make a real and practical use of all the OS-

level improvements described in the previous section,

it is crucial to export all the available features and tun-

ing points to have them available during the place-

ment.

To reach such a result, the model of the underly-

ing infrastructures has to be enhanced, and the cor-

responding descriptors must be enriched to describe

such an enhancement. As an example, consider Open-

Stack and the Heat Orchestration Template (HOT)

used by the OpenStack orchestration engine (Heat) to

describe the system resources. The level of details

currently described by the template has been deﬁned

focusing on homogenous cloud infrastructures and is

unable to exploit hybrids clouds and modern features

provided at the OS level in each physical node. This

lack concerning representativeness leads to a subopti-

mal use of resources also at higher levels of the stack.

For example, Tacker (Orchestration, 2017), the

NFV orchestrator, that is in charge of managing NFV

component cannot utilize speciﬁc OS features in the

MANO (ETSI, 2014) descriptors because they are not

exported by the Virtualized Infrastructure Manager

(VIM) (aka OpenStack). This could jeopardize the

possibility to satisfy the strict latency requirements of

this kind of applications.

To address this issue, the interfaces among the dif-

ferent layers must be enriched allowing to describe

the QoS requirements regarding the computational ca-

pacity at all levels, as shown in Figure 2.

Within the MANO descriptor of each NFV com-

ponent, we can ﬁnd the deﬁnition of the required Vir-

tual Deployment Units (VDU) needed to deploy that

service. Each of them is characterized by a speciﬁ-

cation of the number of required CPUs, which should

be extended to include the capability to declare CPU

fractions deﬁned by budget and period or by utiliza-

tion and maximum service delay, in order to exploit

reservation-based real-time scheduling features. This

enriched VDU descriptors would be processed by the

CLOSER 2018 - 8th International Conference on Cloud Computing and Services Science

630

VIM to produce the corresponding orchestration tem-

plates (HOT), equivalently enhanced, required by the

Heat orchestration engine. With such an approach,

a real-time container or VM could be deployed ac-

cording with the speciﬁed scheduling parameters, af-

ter proper admission control done at the orchestration

layer, exploiting the kernel extensions and modern

scheduling algorithms, such as SCHED DEADLINE.

4.3 Guest Architecture

Although most of the previous discussion focused on

the issues in mapping high-level abstractions to the

host OS and VMM conﬁgurations, similar opportuni-

ties for improvements can be encountered in the con-

ﬁguration of the guest system.

For example, many of the currently used cloud

management stacks end up installing a complete

general-purpose OS (that can run in containers or

in more isolated VMs) for every cloudiﬁed service.

However, better performance (with smaller resource

requirements) can be achieved by properly conﬁgur-

ing the guest OS to take advantage of the virtualiza-

tion features or even using specialized guest OSes. Of

course, this requires some visibility of the virtualiza-

tion infrastructure from the guest.

Although Unikernels (Madhavapeddy et al., 2013;

Kantee, 2015) have been proposed as a way to easily

specialize the guest OS for speciﬁc services and spe-

ciﬁc virtualization technologies, many fo the of the

most commonly used cloud management stacks are

not able to properly exploit this possibility.

Again, an appropriate mapping of abstract service

descriptions to guest system details (and a better de-

scription of the host OS / hypervisor capabilities) is

needed to allow the automatic conﬁguration and de-

ployment of virtualized services with better perfor-

mance.

5 CONCLUSIONS

A major drawback of existing research efforts in per-

formance control and isolation of multiple VMs when

co-located on the same physical hosts resides in their

lack of direct and easy applicability in the context of

real industrial cloud providers. For example, with a

few exceptions, the above works provide prototype

implementations in the form of hacks to the original

open-source software ensemble (e.g., Xen, Linux ker-

nel, KVM sources), rather than well-engineered so-

lutions that can readily be leveraged in higher-level

infrastructure management stacks.

Indeed, industrialization of prototypes as de-

scribed in a research paper can be something totally

non-trivial. As it was the case for SCHED DEADLINE,

having a prototype non-standard mechanism land-

ing into the mainline Linux kernel can cost several

months of development effort, with continuous in-

teractions with core kernel developers and maintain-

ers. Therefore, we have several of these custom ap-

proaches, patches and hacks, that are rarely made

available and released in the public in a usable state,

with the natural side-effect that it is overly difﬁcult

to design higher-level cloud and NFV orchestration

layers that can rely on such experimental low-level

features. On the side of OS vendors and maintain-

ers, it is always difﬁcult to incorporate novel non-

standard features, without a clear idea of the need for

said features by at least a community of users. This

chicken-and-egg situation impairs availability of ad-

vanced performance control features in orchestration

engines, which, coupled with the lack of exposure of

the plethora of capabilities and performance tunables

already available in nowadays OS and hypervisors,

leads to the consequence that it is quite difﬁcult for

real cloud and NFV providers to leverage solid and

important research results in these areas, as well as

existing advanced performance control mechanisms.

In our opinion, an important step in this regard

is to increase exposure of low-level performance tun-

ables to the high-level layers of the infrastructure or-

chestration suite, in order to allow for a proper match-

making process that considers also high-level applica-

tion/service requirements. Therefore, we are working

on further research along these lines.

REFERENCES

Abeni, L., Kiraly, C., Li, N., and Bianco, A. (2015). On the

performance of kvm-based virtual routers. Computer

Communications, 70(Supplement C):40 – 53.

Aesberg, M., Forsberg, N., Nolte, T., and Kato, S. (2011).

Towards real-time scheduling of virtual machines

without kernel modiﬁcations. In ETFA2011, pages 1–

Alicherry, M. and Lakshman, T. V. (2012). Network aware

resource allocation in distributed clouds. In 2012 Pro-

ceedings IEEE INFOCOM, pages 963–971.

Balsini, A., Parri, A., and Abeni, L. (2017). RT bandwidth

constraints enforced by hierarchical DL scheduling.

https://lkml.org/lkml/2017/3/31/658.

Barham, P., Dragovic, B., Fraser, K., Hand, S., Harris, T.,

Ho, A., Neugebauer, R., Pratt, I., and Warﬁeld, A.

(2003). Xen and the art of virtualization. In Proceed-

ings of the Nineteenth ACM Symposium on Operating

Systems Principles, SOSP ’03, pages 164–177, New

York, NY, USA. ACM.

The Importance of Being OS-aware

631

Barr, J. (2015). AWS News Blog – EC2 Instance

History. https://aws.amazon.com/blogs/aws/ec2-

instance-history/.

Brown, K. (2016). Beyond buzzwords: A

brief history of microservices patterns.

https://www.ibm.com/developerworks/cloud/library/cl-

evolution-microservices-patterns/index.html.

Caulﬁeld, A. M., Chung, E. S., Putnam, A., Angepat, H.,

Fowers, J., Haselman, M., Heil, S., Humphrey, M.,

Kaur, P., Kim, J. Y., Lo, D., Massengill, T., Ovtcharov,

K., Papamichael, M., Woods, L., Lanka, S., Chiou,

D., and Burger, D. (2016). A cloud-scale acceleration

architecture. In 2016 49th Annual IEEE/ACM Inter-

national Symposium on Microarchitecture (MICRO),

pages 1–13.

Chang, F., Ren, J., and Viswanathan, R. (2010). Optimal

resource allocation in clouds. In 2010 IEEE 3rd In-

ternational Conference on Cloud Computing, pages

418–425.

Checconi, F., Cucinotta, T., Faggioli, D., and Lipari, G.

(2009). Hierarchical Multiprocessor CPU Reserva-

tions for the Linux Kernel. In Proceedings of the

5th International Workshop on Operating Systems

Platforms for Embedded Real-Time Applications (OS-

PERT 2009).

Cherkasova, L., Gupta, D., and Vahdat, A. (2007). Compar-

ison of the three cpu schedulers in xen. SIGMETRICS

Perform. Eval. Rev., 35(2):42–51.

Corbet, J. (2016). Controlling access to the memory cache.

https://lwn.net/Articles/694800/.

Cortez, E., Bonde, A., Muzio, A., Russinovich, M., Fon-

toura, M., and Bianchini, R. (2017). Resource cen-

tral: Understanding and predicting workloads for im-

proved resource management in large cloud platforms.

In Proceedings of the 26th Symposium on Operating

Systems Principles, SOSP ’17, pages 153–167, New

York, NY, USA. ACM.

Cucinotta, T., Checconi, F., and Giani, D. (2011). Improv-

ing Responsiveness for Virtualized Networking Under

Intensive Computing Workloads. In Proceedings of

the 13th Real-Time Linux Workshop, RTLSWS.

Cucinotta, T., Checconi, F., Kousiouris, G., Kyriazis, D.,

Varvarigou, T., Mazzetti, A., Zlatev, Z., Papay, J.,

Boniface, M., en Berger, S., Lamp, D., Voith, T.,

and Stein, M. (2010). Virtualised e-Learning with

Real-Time Guarantees on the IRMOS Platform. In

Proceedings of the IEEE International Conference on

Service-Oriented Computing and Applications (SOCA

2010), pages 1–8, Perth, Australia.

da Silva, R. A. C. and da Fonseca, N. L. S. (2016).

Topology-aware virtual machine placement in data

centers. Journal of Grid Computing, 14(1):75–90.

ETSI, N. (2014). Network functions virtuali-

sation (NFV); management and orchestra-

tion. http://www.etsi.org/deliver/etsi gs/NFV-

MAN/001 099/001/01.01.01 60/gs nfv-

man001v010101p.pdf.

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

real-time virtual resources. In 23rd IEEE Real-Time

Systems Symposium, 2002. RTSS 2002., pages 26–35.

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.

Gai, K., Qiu, M., Zhao, H., Tao, L., and Zong, Z. (2016).

Dynamic energy-aware cloudlet-based mobile cloud

computing model for green computing. J. Netw. Com-

put. Appl., 59(C):46–54.

Garﬁnkel, S. (2011). MIT Technology Review –

Intelligent Machines – Computing Pioneer Dies.

https://www.technologyreview.com/s/425913/

computing-pioneer-dies/.

Ghribi, C., Hadji, M., and Zeghlache, D. (2013). En-

ergy efﬁcient vm scheduling for cloud data centers:

Exact allocation and migration algorithms. In 2013

13th IEEE/ACM International Symposium on Cluster,

Cloud, and Grid Computing, pages 671–678.

Information Resources Management Association (2012).

Grid and Cloud Computing: Concepts, Methodolo-

gies, Tools and Applications. IGI Global, Hershey,

PA, USA.

Kantee, A. (2015). The rise and fall of the operating system.

USENIX login, 40(5).

Kim, H., Lim, H., Jeong, J., Jo, H., and Lee, J. (2009). Task-

aware virtual machine scheduling for i/o performance.

In Proceedings of the 2009 ACM SIGPLAN/SIGOPS

International Conference on Virtual Execution Envi-

ronments, VEE ’09, pages 101–110, New York, NY,

USA. ACM.

Kim, H. and Rajkumar, R. (2016). Real-time cache manage-

ment for multi-core virtualization. In 2016 Interna-

tional Conference on Embedded Software (EMSOFT),

pages 1–10.

Konstanteli, K., Cucinotta, T., Psychas, K., and Varvarigou,

T. A. (2014). Elastic admission control for federated

cloud services. IEEE Transactions on Cloud Comput-

ing, 2(3):348–361.

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

Deadline scheduling in the linux kernel. Software:

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

Madhavapeddy, A., Mortier, R., Rotsos, C., Scott, D.,

Singh, B., Gazagnaire, T., Smith, S., Hand, S., and

Crowcroft, J. (2013). Unikernels: Library operating

systems for the cloud. SIGARCH Comput. Archit.

News, 41(1):461–472.

Mann, Z. A. (2015). Allocation of virtual machines in

cloud data centers—a survey of problem mod-

els and optimization algorithms. ACM Comput. Surv.,

48(1):11:1–11:34.

Networks, J. (2017). Network transforma-

tion with NFV and SDN – a journey to-

ward sustainable competitive advantage.

http://www.juniper.net/assets/us/en/local/pdf/

whitepapers/2000628-en.pdf.

Orchestration, T. O. (2017).

https://wiki.openstack.org/wiki/Tacker. Last ac-

cessed on Dec 6, 2017.

Pellerin, D. (2017). Accelerated Computing on

AWS – Applications for GPUs and FPGAs.

www.asapconference.org/slides/amazon.pdf.

Pongrcz, G., Molnr, L., and Kis, Z. L. (2013). Removing

roadblocks from sdn: Openﬂow software switch per-

formance on intel dpdk. In 2013 Second European

CLOSER 2018 - 8th International Conference on Cloud Computing and Services Science

632

Workshop on Software Deﬁned Networks, pages 62–

67.

Public, C. (2016). Align operations for net-

work functions virtualization environments.

https://www.cisco.com/c/dam/en/us/solutions/

collateral/service-provider/network-functions-

virtualization-nfv/white-paper-c11-737101.pdf.

Rizzo, L. (2012a). netmap: A novel framework for fast

packet i/o. In 2012 USENIX Annual Technical Con-

ference (USENIX ATC 12), pages 101–112, Boston,

MA. USENIX Association.

Rizzo, L. (2012b). Revisiting network i/o apis: The netmap

framework. Queue, 10(1):30:30–30:39.

Rizzo, L. and Lettieri, G. (2012). Vale, a switched ethernet

for virtual machines. In Proceedings of the 8th Inter-

national Conference on Emerging Networking Exper-

iments and Technologies, CoNEXT ’12, pages 61–72,

New York, NY, USA. ACM.

Satyanarayanan, M., Bahl, P., Caceres, R., and Davies, N.

(2009). The case for vm-based cloudlets in mobile

computing. IEEE Pervasive Computing, 8(4):14–23.

SCTE (2016). Scte analysis of available energy 2020 par-

ticipating mso data. Brochure.

Steiner, M., Gaglianello, B. G., Gurbani, V., Hilt, V.,

Roome, W., Scharf, M., and Voith, T. (2012).

Network-aware service placement in a distributed

cloud environment. SIGCOMM Comput. Commun.

Rev., 42(4):73–74.

Valcarenghi, L., Giannone, F., Manicone, D., and Castoldi,

P. (2017). Virtualized enb latency limits. In 2017 19th

International Conference on Transparent Optical Net-

works (ICTON), pages 1–4.

Vitucci, C. and Larsson, A. (2017). Flexible 5G Edge Server

for Multi Industry Service Network. International

Journal on Advances in Networks and Services, 10(3-

4). ISSN: 1942-2644.

White, R. (2015). Introduction to Virtualization.

http://www.vm.ibm.com/devpages/bitner/

presentations/virtualb.pdf.

Xi, S., Li, C., Lu, C., Gill, C. D., Xu, M., Phan, L. T. X.,

Lee, I., and Sokolsky, O. (2015). Rt-open stack: Cpu

resource management for real-time cloud computing.

In 2015 IEEE 8th International Conference on Cloud

Computing, pages 179–186.

Xi, S., Wilson, J., Lu, C., and Gill, C. (2011). Rt-xen: To-

wards real-time hypervisor scheduling in xen. In 2011

Proceedings of the Ninth ACM International Confer-

ence on Embedded Software (EMSOFT), pages 39–48.

Xia, Y., Zhou, M., Luo, X., Pang, S., and Zhu, Q.

(2015). Stochastic modeling and performance analy-

sis of migration-enabled and error-prone clouds. IEEE

Transactions on Industrial Informatics, 11(2):495–

504.

Yao, Y., Cao, J., and Li, M. (2013). A network-aware vir-

tual machine allocation in cloud datacenter. In Pro-

ceedings of the 10th IFIP International Conference on

Network and Parallel Computing - Volume 8147, NPC

2013, pages 71–82, New York, NY, USA. Springer-

Verlag New York, Inc.

Yun, H., Yao, G., Pellizzoni, R., Caccamo, M., and Sha, L.

(2013). Memguard: Memory bandwidth reservation

system for efﬁcient performance isolation in multi-

core platforms. In 2013 IEEE 19th Real-Time and

Embedded Technology and Applications Symposium

(RTAS), pages 55–64.

Zhang, J., Wang, M., Luo, J., Dong, F., and Zhang,

J. (2015). Towards optimized scheduling for data-

intensive scientiﬁc workﬂow in multiple datacenter

environment. Concurr. Comput. : Pract. Exper.,

27(18):5606–5622.

The Importance of Being OS-aware

633