Proactive Workload Management for Bare Metal
Deployment on Microservers
Daniel Schlitt
1
, Christian Pieper
1
and Wolfgang Nebel
2
1
OFFIS - Institute for Information Technology, Escherweg 2, 26121 Oldenburg, Germany
2
Carl von Ossietzky Universit
¨
at Oldenburg, Ammerl
¨
ander Heerstraße 114-118, 26111 Oldenburg, Germany
Keywords:
Bare Metal, Data Center, Energy Efficiency, Forecast, Microserver, OpenStack, Workload Management.
Abstract:
This paper introduces a concept for an energy-aware workload management (WM) for heterogeneous mi-
croserver environments. Its main focus is on highly dynamic service-driven workloads often coupled to user
requests requiring fast response times. The WM is developed in scope of the M2DC (Modular Microserver
Data Center) project, in which a new server generation of composable microservers is designed. Targeting
an easy industrial applicability, the underlying middleware is based on a turnkey OpenStack platform. As
part of that middleware, the WM makes use of workload/utilization and power data as well as correspond-
ing (prediction) models to deploy applications on the most suitable microservers and temporarily shut down
unused capacities, either proactively or reactively (in case of deviations from forecasts). The WM has been
implemented and simulated within a virtual environment. However, the integration, refinement and evaluation
on the new M2DC hardware is still work in progress.
1 INTRODUCTION
Technically, the workload handled by computer sys-
tems can be separated into two rough types. First,
there are high performance workloads or batch jobs
which are placed once and just need to be completed
in a certain time frame. Besides, there are transactio-
nal or service-driven workloads, which usually con-
sist of smaller requests placed by users expecting a
timely response. In terms of workload management
(WM), the first mentioned domain benefits from pop-
ular management tools like SLURM, while the lat-
ter have either limited support or functionality es-
pecially in commercial off-the-shelf (COTS) manage-
ment SW.
Nonetheless, dynamic management of transac-
tional workloads holds high potentials. The tight user
coupling makes a both effective and efficient static
operation complicated, as the workload and user be-
havior may be highly dynamic. That means the hard-
ware capacity has to be specified for the maximum
user demand based on experience and estimations.
Additionally, low response times are highly important
from the user’s point of view. These factors make it
impossible to find an optimal static allocation, as the
dynamic workload behavior has to be considered for
compute node selection and application deployment.
In contrast to high performance data centers, which
try to fully utilize all of its compute nodes at any time,
in the case of service-based computing the hardware
usage has to be adapted to actual user demand. By dy-
namically adapting the number and selection of active
nodes, the energy efficiency of data center operation
can be massively increased.
Therefore, we realize an energy-aware WM based
on a two-folded approach: (1) we optimize the cur-
rent state of operation (active compute nodes, appli-
cation allocation) based on workload/utilization pre-
dictions as well as performance/power models and
(2) in the case of non-matching forecasts we use a
reactive allocation to adjust the amount of deployed
applications, while the deviating prediction models
are re-trained. This currently ongoing work is done
in scope of the EU H2020 project M2DC
1
(Modular
Microserver Data Center), which aims to provide full
server appliances with a convincing total cost of own-
ership (TCO). The project is based on three pillars
(Oleksiak et al., 2016): It offers freely configurable
heterogeneous microservers of different architectures
(x86/64, ARM64) and hardware accelerators (GPU,
FPGA). Furthermore, there is a layer of advanced
management strategies e.g. to optimize the overall en-
1
http://www.m2dc.eu
246
Schlitt, D., Pieper, C. and Nebel, W.
Proactive Workload Management for Bare Metal Deployment on Microservers.
DOI: 10.5220/0006762202460253
In Proceedings of the 7th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2018), pages 246-253
ISBN: 978-989-758-292-9
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
Intelligent Power
Management
Photo HPC Cloud Your
Data Centre
integration
Figure 1: M2DC project scheme.
ergy efficiency. The described WM within this paper
would be part of that layer. In order to enable an easy
industrial applicability, M2DC is designed as turnkey
platform with well-defined interfaces to the surround-
ing ecosystem (see Figure 1). For that purpose, the
middleware is based on OpenStack, which is an open
source software project offering a free cloud comput-
ing architecture. It is backed up by a large commu-
nity, which steadily extends OpenStack’s function-
alities. New functions and features are made avail-
able through components, which implement e.g. sup-
port for networking infrastructure (Neutron), work-
load scheduling (Nova), management of bare metal
resources (Ironic) or a web interface (Horizon). The
WM, introduced in this paper, is part of the middle-
ware and has the objective to dynamically optimize
the server operation to contribute to the overall TCO
reduction. As OpenStack including the Ironic
2
com-
ponent – its bare metal provisioning module – is used
as a base platform for the middleware, the WM is con-
ceptionalized as an add-on to OpenStack.
In this paper, we describe the conception of the
WM. Following the Introduction, Section 2 gives an
overview on known approaches for WM in data cen-
ters. After that, we present our current development
stage. The implementation is done in two separate
places. On the one hand, we created an intelligent
management component containing required estima-
tion models and forecast algorithms. This component
2
https://docs.openstack.org/ironic/latest/
initiates the proactive and reactive node management
and is able to access our OpenStack Nova
3
exten-
sion which was implemented on the other hand. As
Nova is responsible for the actual node scheduling,
we introduce new filters for an improved server se-
lection based on energy efficiency and performance.
Accordingly, Section 3 summarizes all relevant data
sources and models used by the allocation algorithm.
The general structure of the intelligent management
including workload management, the allocation algo-
rithm and the scheduling process will be explained in
Section 4. Afterwards, Section 5 shows preliminary
simulation results before Section 6 presents the con-
clusion and an outlook to the near future.
2 STATE OF THE ART
In research, there are dozens of approaches and im-
plementations for server WM in data centers. Jen-
nings and Stadler did an exhaustive literature sur-
vey and present their results in (Jennings and Stadler,
2015). We will focus on related work with respect to
the most striking aspects of M2DC’s WM, being (1)
management of heterogeneous hardware nodes, (2)
proactive management by using workload forecasts
and server models and (3) setting up on OpenStack
as most widespread cloud management system.
3
https://docs.openstack.org/nova/latest/
Proactive Workload Management for Bare Metal Deployment on Microservers
247
Management of Heterogeneous Hardware. Jen-
nings and Stadler (Jennings and Stadler, 2015) present
an exhaustive list of WM approaches using dynamic
virtualization. Some of them have also been ap-
plied to industrial products. A good example is the
distributed resource scheduling (DRS) and its exten-
sion distributed power management (DPM) feature
by VMware for its virtualization technology, offer-
ing significant energy savings by dynamic VM mi-
gration and minimization of physical machines (Gu-
lati et al., 2012). Applying such solutions requires
fully virtualized data centers, but virtualization tech-
nology is currently limited to x86/64 and ARM64
and it has some considerable overhead on small mi-
croserver nodes. Therefore, it is not the best solution
for M2DC with its heterogeneous hardware philos-
ophy, supporting a wide spectrum of different com-
pute nodes ranging from low power to high perfor-
mance. Still, there are some investigations using other
technologies for WM. Piraghaj et al. (Piraghaj et al.,
2015) as well as Kang et al. (Kang et al., 2017) pro-
pose WM based on Docker containers to additionally
reduce the virtualization overhead, which would also
be suitable for low-power microservers. De Assunc¸ao
et al. (de Assunc¸ao et al., 2016) implement dynamic
server provisioning using bare metal provisioning via
OpenStack Nova and an extension of its built-in fil-
ter and weighing scheduler. However, these alterna-
tive approaches still do not include accelerators like
FPGAs or GPUs.
Proactive Workload Management. Some ap-
proaches consider workload forecasts in order to
achieve better consolidation rates with less buffer ca-
pacities. Zhang et al. (Zhang et al., 2014) propose an
algorithm which schedules work in a hybrid cloud be-
tween on- and off-premises compute capacities. On-
premises in terms of software or workload which
is processed on company owned or rented servers,
whereas off-premises means workload processing on
remote facilities like cloud computing or Software as
a Service (SaaS). The researcher divide workload in
’base’ load and ’flash crowd’ load, which are man-
aged individually. Base load is managed proactively
on-premises (90 % of time within 17 % prediction er-
ror) and flash crowd load is managed reactively off-
premises. Currently, their approach does not allow for
dynamic workload scheduling and it is restricted to
certain workload types, which impedes an automatic
usage on unknown applications. Herbst et al. (Herbst
et al., 2014) use time series analysis to forecast work-
loads and use this information for proactive resource
scheduling via virtualization. They present an ex-
haustive comparison of diverse statistical methods.
The most suitable time series model is selected at run
time based on the given context. By dynamically se-
lecting the model, the relative error could be reduced
by 37 % on average compared to statically selected
fixed forecasting methods. A similar approach is uti-
lized in M2DC.
Cloud Management based Realization. The
adaptability in real, productive data centers is an
important criterion for the success of novel WM
methods. The best approach would be to set up
on broadly known and accepted server/cloud man-
agement tools to have a potential user base with
fewer contraints compared to proprietary approaches.
There is only little work in research, which already
use common management tools as a base for imple-
menting their own solutions. The most promising
one is from Beloglazov and Buyya (Beloglazov and
Buyya, 2015), who use OpenStack for their dynamic
consolidation module called Neat. It is designed
as a transparent add-on not requiring modifications
on OpenStack installations. Neat makes use of
OpenStack’s management functions via public APIs
to perform the VM migrations planned by the Neat
allocation algorithm. However, Neat is restricted
to virtualization and therefore not the optimal so-
lution for M2DC’s smaller microservers. Fujitsu
is also working on a management tool called FU-
JITSU Software ServerView Resource Orchestrator
(ROR) (Yanagawa, 2015) building up on OpenStack.
In contrast to Neat, autonomic functions for WM are
missing, as the focus rather is on increasing the ease
of operation to make management functions usable
for business purposes. Then again it is planned to
integrate OpenStack Ironic in the future to support
bare metal deployment next to virtualization.
M2DC’s energy-aware WM combines the advan-
tages of several referenced approaches which each
focus only on limited aspects. By applying a com-
bined proactive and reactive allocation algorithm, the
M2DC WM is able to optimize aggressively while
also providing emergency measures for sudden spikes
in workload. However, the most appealing advantage
of M2DC’s approach is the strict focus on future ap-
plicability. While the usage of OpenStack as base
platform should help spreading the solution due to its
broad community and popularity, the consideration of
alternative compute nodes in the modeling and man-
agement process guarantees a future relevance when
GPUs and FPGAs become more popular in general-
purpose computing.
SMARTGREENS 2018 - 7th International Conference on Smart Cities and Green ICT Systems
248
3 MODELS
The WM improves the overall application-to-node al-
locations. For the decision making it requires data
sources of directly measured operational parameters
as well as models, which make estimations based on
collected data. This section briefly introduces re-
quired parameters, but does not go into details. It is
rather an overview for further WM explanations.
Data Sources. Data sources are separated in ve
parts: The workload is the amount of work a com-
pute node has to tackle. Workload, in the form of
jobs, may either be placed once for direct processing
or may come over time based on current demand or
user requests. Performance is defined by the work-
load processed per time for a given compute node.
That means performance is also application specific
and can therefore only be measured for adapted appli-
cations. Benchmarks help to measure hardware per-
formance for application workloads in order to com-
pare capabilities of compute nodes. This must be
done prior to productive operations. Utilization rep-
resents the usage of a compute node’s subsystems by
the applied workload and can be used as objectively
determinable proxy for workload data, when access
to application data is not available. The power will
be measured for each compute node, e.g. to deter-
mine the energy efficiency. In conjunction with re-
source utilization data, power models can be used to
estimate power demand for arbitrary node utilization.
The thermal and resource management aggregates the
thermal state of compute nodes in a thermal met-
ric. For this purpose it scores available temperature
sensors or temperature affecting components like fan
speed into a ratio, which can be used for later place-
ment decisions.
Models. Using the collected data sources, mod-
els are able to make calculations for defined time
points, even in the future. The WM requires models
for power, performance, energy efficiency and work-
load/utilization. Power models are able to calculate
the power demand based on historic and future work-
load or utilization trends. Power data is required
for other models like the energy efficiency. Perfor-
mance models determine the performance of M2DC
components. In general, the objective is to use pub-
lished benchmark results in order to avoid measure-
ments on productive systems, cf. (Schlitt and Nebel,
2016). As the server benchmarks cannot be applied
to every hardware type of M2DC compute nodes, an
application-specific method will be required. There-
fore use case applications must be designed con-
sidering external monitoring of performance factors.
The energy efficiency model is defined as relation be-
tween performance and the amount of required en-
ergy. Workload modeling is a main selling proposition
of M2DC’s WM. It utilizes workload or utilization
forecasts for the scheduling process, taking into ac-
count the user behavior to enable proactive actions for
regular peaks. For this, the WM uses a number of uni-
variate, statistical procedures to automatically model
unknown, generic time series solely based on historic
values. These are Seasonal and Trend decomposition
using Loess (STL), Holt-Winters (HW) filtering and
Seasonal AutoRegressive Integrated moving Average
(SARIMA), which results are compared to choose the
best approach for the specific situation.
Storage. Measured values are stored in a central
database using Gnocchi
4
of the underlying Open-
Stack, while models required for the estimation of
values are placed within the intelligent management.
4 WORKLOAD MANAGEMENT
The WM component is embedded in the intelligent
management (IM) layer of M2DC’s middleware. The
IM layer is depicted in Figure 2. The central com-
ponent is the IM component, which acts as an in-
terface combining dataflows and workflows between
M2DC’s management components, the (model) data
and OpenStack.
The IM is the connection between the resource
and thermal management (RTM), the Power Man-
agement (PM) and the WM components. There are
two different PM solutions, one of them switching
host states by using the OpenStack API and the other
using directly the server controller API in case fur-
ther power management functionality is needed (e.g.
power capping) or if usage of OpenStack and WM is
not desired. Basically, WM and RTM could have dif-
ferent constraints, which have to be negotiated before
instructions are sent to the PM. Access to the server
models (power, performance, ...), thermal and work-
load models as well as measured data in the Gnocchi
database is realized through references to the com-
ponents, gathered in the IM class. Thereby, RTM
and WM can access each kind of model and use ev-
ery information available for management decisions.
Moreover, the IM represents the interface to Open-
Stack Nova regarding the dynamic scheduling. While
the allocation algorithm itself is part of the WM class,
connection to the Compute API as well as the filter
4
http://gnocchi.xyz/
Proactive Workload Management for Bare Metal Deployment on Microservers
249
Efficient Management
OpenStack
Models
Management
Support
Thermal and
Resource
Management
Server Model
Storage
Workload
Forecasting
Intelligent
Management
Power
Management
Gnocchi API
Compute API
M2DC
BaseHostFilter
Workload
Management
Thermal Filter
Perf Filter
EE Filter
Figure 2: Overview on intelligent management component.
classes are established via IM. The Compute API is
used for triggering application deployment and set-
ting host power states.
The WM’s main objective is to optimize the
application-to-node allocation in order to increase the
overall energy efficiency, i.e. the WM tries to se-
lect the most suitable compute nodes for the applica-
tions regarding current and future workload demands
e.g. using low power ARM nodes if utilization is
low at night. Therefore, the current and future state
is analyzed and corresponding actions are triggered
by an allocation algorithm. Suitable actions are ex-
ecuted by the corresponding components (e.g. power
management for host power actions), which are ac-
cessed via IM. Moreover, the WM provides general
methods for creating, deploying, moving and destroy-
ing application instances, which are called, when the
user takes the corresponding actions in OpenStack.
Also, the WM has means to read the current state
of hosts/compute nodes and instances inside Open-
Stack, which are analyzed by the allocation algorithm.
These methods ensure that the allocation states be-
tween WM and OpenStack are consistent.
The allocation algorithm is the centerpiece of
WM. Generally, it distinguishes between two opera-
tional states: proactive and reactive allocation. The
operational state is constantly determined by compar-
ing current, actually measured utilization/workload
with forecast values. If the deviation is within a tol-
erance range (e.g. 20 %), allocation optimization is
done proactively based on forecasts. If the actual
utilization is below the threshold, the allocation will
still be optimized proactively except that the workload
modeling process for the affected compute nodes will
be repeated beforehand (maximum once per day). In
case the tolerance thresholds and the node utilization
thresholds (e.g. 80 %) are exceeded, allocation opti-
mization is interrupted and a high priority reactive
allocation will be performed. Afterwards, workload
models of concerning nodes will be reprocessed and
the proactive allocation optimization continues.
In proactive operation, the allocation algo-
rithm analyzes the current and future host utiliza-
tions/workloads and decides for each node (1) if ad-
ditional compute capacities have to be provided on
short-term, (2) the state may be optimized or (3) if it
should remain constant. As booting and deployment
activities take several minutes, a time window of at
least 30 minutes should be chosen for workload anal-
yses. Then again, the window must not be too long, as
that would reduce the optimization potential because
the allocation has to be valid (not exceeding utiliza-
tion thresholds) for the whole time window.
The algorithm starts to examine the provisioning
of more compute capacities. This is necessary, if the
expected utilization of a single node will exceed its
utilization threshold (e.g. 80 %) in the investigated
time window. This will initiate the deployment of the
affected application on another node, as every node is
expected to maintain a certain predetermined buffer.
The scheduling of a new node will be triggered and
executed in time to ensure the necessary capacities are
available as soon as they are needed. Subsequently,
the original node will be powered down, if the new
node has enough computing capacities. After provi-
sioning additional capacities, the utilization trend is
analyzed for the new allocation and the process is re-
peated, if necessary.
When all cases for capacity provisioning have
been resolved, the algorithm looks into potential en-
ergy efficiency optimization, which is often achieved
by utilization decreasing over time (e.g. after-work
SMARTGREENS 2018 - 7th International Conference on Smart Cities and Green ICT Systems
250
hours, night). The algorithm starts with the currently
least energy efficient node (determined by energy ef-
ficiency model) and looks (1) for potential consolida-
tion of the running application instance or (2) for po-
tential alternative nodes capable of running the same
application more efficiently. Instance consolidation is
possible, if all other nodes running the same appli-
cation are able to provide enough capacities to take
over the workload of the instance which will be shut
down. Thus, an instance and node shutdown will be
triggered and called in OpenStack Nova via IM and
the Compute API. If this approach is not feasible, the
algorithm checks the availability of a more energy ef-
ficient node for given utilization. If energy would
be saved in the investigated interval despite deploy-
ment overhead and buffer utilization is still met, the
scheduling as well as a subsequent shutdown is trig-
gered for the selected node/application. Anyway, the
next node is examined in the same way until every
node was examined. The proactive allocation process
will be restarted when half of the investigated time
window is over, thereby ensuring that future utiliza-
tion increases can be handled in time.
A reactive allocation is done for each compute
node, whose actual utilization exceeds its forecast by
the predetermined threshold. Thereby, the scheduling
of an additional compute node for the affected appli-
cation will be triggered before doing anything else, as
one can no longer rely on the utilization forecasts and
it is not known how the current trend will continue.
Moreover, the scheduling will be configured to pick
preferably high performance nodes in order to coun-
tervail a further increasing utilization. After all nec-
essary reactive operations have been performed and
workload modeling has been rerun, the allocation al-
gorithm tries to proactively optimize the current state.
All mentioned configuration parameters (thresh-
olds, percentages, times) are preliminary and have to
be adapted based on empirical evaluations.
4.1 Scheduling via Nova
The actual scheduling of applications to compute
nodes is done via the OpenStack Nova FilterSched-
uler
5
, using several host filters and weighers. This
scheduler is extended by an M2DCBaseHostFilter,
which is derived from the Nova BaseHostFilter. It
has three child classes representing the considered as-
pects in WM and RTM. With regard to WM there is an
energy efficiency and a performance filter. The energy
efficiency filter passes through only nodes/hosts that
have at least the minimum specified energy efficiency
5
https://docs.openstack.org/nova/latest/user/filter-
scheduler.html
threshold, which was assigned by WM. The utiliza-
tion dependent node energy efficiency is computed
by using the models suggested in Section 3, which
are accessed via IM. The performance filter works
analogous. Regarding RTM, there is a thermal filter
passing through nodes/hosts based on a thermal met-
ric, which indicates if the corresponding node should
be considered for deployment or if the thermal en-
vironment is currently not suitable (e.g. thermal hot
spots). After the filtering process, the most suitable
host is selected by weighed sorting of all passed com-
pute nodes by using adapted M2DCBaseHostWeigher.
The weights are again given by the WM and depend
generally on expected utilization/workload levels, e.g.
if the expected utilization is low, the more energy ef-
ficient nodes are preferred and would benefit from a
higher weighing.
As the allocation algorithm examines the opera-
tional state of compute nodes and instances, it trig-
gers scheduling processes for a given application with
state-specific parameters defining energy efficiency,
performance and thermal requirements. In case there
is no suitable host for the triggered scheduling process
(i.e. no host passed all filters), the allocation algorithm
will get an appropriate response and it will ignore this
node as optimization target until the next power man-
agement operation was resolved. If the scheduling
was triggered due to reactive allocation, the admin-
istrators will be notified that there is no suitable host
to provide the needed capacity.
4.2 Microserver Support
M2DC’s main unique selling point is the support
of a high variety of different hardware architectures
in the microserver format. Within M2DC, Open-
Stack is extended to support dynamic composition
and provisioning of diverse microserver nodes includ-
ing FPGAs and GPUs. As WM is part of the base
M2DC installment, all available microserver types
have to be supported. This mainly influences whether
an application is deployable on a certain microserver
or not, depending on available implementations and
(boot) images.
The allocation algorithm itself does not consider
whether an application can be deployed on a certain
microserver. This aspect is checked in the destination
host selection of OpenStack Nova filter scheduler via
the ImagePropertiesFilter. This filter only passes
hosts satisfying the requirements given with the ap-
plication image to deploy. If multiple images (e.g. for
different architectures) are available, several requests
will be sent consecutively.
Proactive Workload Management for Bare Metal Deployment on Microservers
251
Table 1: First simulation results for static vs. dynamic (optimized) allocation.
Scenario Configuration Energy (static) Energy (optimized) Savings
1 2 x86 (2013) + 2 ARM64 (2015/16) 3372 Wh 2082 Wh 38.2%
2 4 x86 (2013) 3234 Wh 2382 Wh 26.2%
3 2 x86 (2017) 1710 Wh 1710 Wh 0.0%
4 4 ARM64 (2016) 1860 Wh 1440 Wh 22.5%
4.3 Power Management
The power management is an essential part of
M2DC’s IM in order to control the overall energy
demand of the system. Realized as support compo-
nent, it encapsulates access to different power actions,
which comprise changes to power states but also
functions like power capping or frequency scaling.
M2DC’s power management components use two dif-
ferent approaches. On the one hand, OpenStack needs
to control nodes for its scheduling process, where e.g.
hosts need to be turned on prior to deploying appli-
cation instances. For the M2DC server, extensions
will take care to implement host-actions function-
ality of Nova, respectively node-set-power-states of
Ironic, using the provided M2DC server API. On the
other hand, it is necessary to execute basic PM func-
tions without using the OpenStack platform. There-
fore, there is an additional power management imple-
mentation that accesses the M2DC hardware directly.
These functions include fan management, power cap-
ping or frequency scaling demanded by RTM, which
are accessed via M2DC server API or directly on the
microserver.
5 PRELIMINARY RESULTS
So far, the IM component, the WM algorithms and the
OpenStack Nova extensions have been implemented
in a virtual environment based on DevStack. To test
the functional capability as well as the effectiveness
of the algorithms, we expanded the DevStack envi-
ronment with a simulation. As simulation data we
modeled several server systems (x86 servers from
2013 and 2017) and microservers (ARM64 from 2015
and 2016) with regard to performance, power and en-
ergy efficiency each dependent on a given utiliza-
tion. The models base to some extent on own mea-
surements but for the most part on publicly available
benchmark results such as SPECpower ssj2008
6
and
are defined by polynomial functions with utilization
as variable. As input data we used utilization forecasts
for two real application workloads measured in a pro-
ductive data center. In our simulation, both applica-
6
https://www.spec.org/power ssj2008/
tions may be run in several instances distributed over
different compute nodes, but not more than one appli-
cation instance per node (bare metal approach). We
simulated several small node configurations each with
a static allocation of application instances to compute
nodes as well as a dynamic workload management.
The simulation was accelerated by factor 60 so that
the run-time of 10 minutes represented 600 minutes
in reality. The results for running each configuration
can be found in Table 1.
In scenario 3, there was no optimization feasible
because the two applications have to run on different
server nodes due to the bare metal approach (opposed
to virtualization, which can potentially operate both
applications on a single node). However, scenarios 1,
2 and 4 show significant savings even in small config-
urations with few optimization possibilities. It is in-
teresting that the mixed configuration in scenario 1 is
more efficient than the homogeneous configuration in
scenario 2, if workload management is applied, while
it needs more energy than scenario 2 in a static opera-
tion. This is an indicator for efficiency gain potential
by using heterogeneous computing nodes. Another
interesting result is the comparison of scenarios 3 and
4. While the modern x86 systems are better in a static
operation, the ARM64 configuration provides a more
fine-granular optimization potential, although this is
somewhat impaired by the small scenario configura-
tions.
We also simulated the productive server environ-
ment of one of M2DC’s partners as of the year 2013
(which is the base year for comparisons in M2DC
project) consisting of 40 x86 server systems. There-
fore, we modeled the server systems from 2013 re-
garding performance, power and energy efficiency.
We also measured the workload of our target appli-
cation in 2016 and used the corresponding forecast
model as input for the simulation. The result was
that the application of WM could reduce the energy
demand by about 800 kWh per month (> 40%) by
simply taking advantage of the high daily variance in
workload profiles.
SMARTGREENS 2018 - 7th International Conference on Smart Cities and Green ICT Systems
252
6 CONCLUSIONS
The M2DC project consortium is working on bring-
ing alternative server technologies like ARM based
processors, GPUs and FPGAs into commercial data
centers. The aim is to decrease capital and opera-
tional expenses by using microservers, which are both
cost- and energy-efficient as they can be specifically
selected and configured for given tasks. As partner in
M2DC, we are specifically working on WM mecha-
nisms which make use of workload/utilization, per-
formance and power data to place applications on
the most suitable microservers and temporarily shut
down unused capacities. A first simulation of the al-
gorithms show promising results of about 40% en-
ergy savings. The related work section shows that
there is already much work in this field. However,
the M2DC approach is the only one (to best of our
knowledge) which combines the advantages of using
workload/utilization forecasts, setting up on a com-
monly known platform (OpenStack) and supporting
alternative (modern) server architectures.
The next steps on the M2DC WM include the
automatic modeling of server power, performance
and energy efficiency models as well as automatic
model selection and modeling of application work-
load/utilization forecasts. After fine tuning and fur-
ther simulations in the virtual environment the IM in-
cluding its interface extensions to OpenStack will be
integrated in the M2DC testbed containing the newly
developed microserver nodes. Evaluations on this
testbed with use cases defined in M2DC will then be
published in a future research paper.
ACKNOWLEDGEMENTS
This scientific work has received funding from the
European Union’s Horizon 2020 research and inno-
vation programme under grant agreement No. 688201
(M2DC).
REFERENCES
Beloglazov, A. and Buyya, R. (2015). Openstack neat: a
framework for dynamic and energy-efficient consoli-
dation of virtual machines in openstack clouds. Con-
currency and Computation: Practice and Experience,
27(5):1310–1333.
de Assunc¸ao, M. D., Lefevre, L., and Rossigneux, F. (2016).
On the impact of advance reservations for energy-
aware provisioning of bare-metal cloud resources. In
CNSM 2016.
Gulati, A., Holler, A., Ji, M., Shanmuganathan, G., Wald-
spurger, C., and Zhu, X. (2012). Vmware distributed
resource management: Design, implementation, and
lessons learned. VMware Technical Journal, 1(1):45–
64.
Herbst, N. R., Huber, N., Kounev, S., and Amrehn, E.
(2014). Self-adaptive workload classification and
forecasting for proactive resource provisioning. Con-
currency and computation: practice and experience,
26(12):2053–2078.
Jennings, B. and Stadler, R. (2015). Resource management
in clouds: Survey and research challenges. Journal of
Network and Systems Management, 23(3):567–619.
Kang, D.-K., Choi, G.-B., Kim, S.-H., Hwang, I.-S., and
Youn, C.-H. (2017). Workload-aware resource man-
agement for energy efficient heterogeneous docker
containers.
Oleksiak, A., Rosinger, S., Schlitt, D., Pieper, C., et al.
(2016). Data centres for iot applications: The m2dc
approach (invited paper). In 2016 International Con-
ference on Embedded Computer Systems: Architec-
tures, Modeling and Simulation (SAMOS), pages 293–
299.
Piraghaj, S. F., Dastjerdi, A. V., Calheiros, R. N., and
Buyya, R. (2015). A framework and algorithm for
energy efficient container consolidation in cloud data
centers. In Data Science and Data Intensive Sys-
tems (DSDIS), 2015 IEEE International Conference
on, pages 368–375. IEEE.
Schlitt, D. and Nebel, W. (2016). Data center perfor-
mance model for evaluating load dependent energy ef-
ficiency. In Proceedings on 4th International Confer-
ence ICT for Sustainability 2016 (ICT4S 2016). At-
lantis Press.
Yanagawa, T. (2015). OpenStack-based Next-generation
Cloud Resource Management. Fujitsu Sci. Tech. J,
51(2):62–65.
Zhang, H., Jiang, G., Yoshihira, K., and Chen, H. (2014).
Proactive workload management in hybrid cloud com-
puting. IEEE Transactions on Network and Service
Management, 1(11):90–100.
Proactive Workload Management for Bare Metal Deployment on Microservers
253