![](bg3.png)
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