1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
|
\chapter{Conclusion}\label{s:conclusion}
Datacenter manageability is a top-priority for the digital society.
Over 3 million jobs in the Netherlands directly depend on cloud services, which are hosted in datacenters~\cite{DBLP:journals/corr/IosupKLVG22}.
Datacenter digital twinning, a promising management technique can offer unique insight into complex warehouse behaviour~\cite{DBLP:journals/computer/AthavaleBBMMPS24}.
In this thesis we paved the way to more advanced \gls{dcdt}s.
We contribute to the scientific community a set of findings that we hope will prove helpful in enabling predictive analytics in both existing \gls{dcdt}s and future projects.
Starting from a thorough investigation into the new, emerging field of datacenter digital twinning, we designed a system capable of incorporating sophisticated data analysis techniques.
We ended our project with a novel evaluation method used in a set of exhaustive experiments.
We answer the main research question by addressing each sub-research question.
\section{Answers to Each Research Question}\label{ss:answers_to_rqs}
\begin{enumerate}[label=\emph{RQ\textsubscript{\arabic*}}, itemsep=1em]
\item \emph{How to asses the current state-of-the-art of digital twinning for datacenters?}\\
To answer this research question, we conducted a semi-structured literature review.
Our findings indicate that the field of datacenter digital twinning is still under development, and there exist few \gls{dcdt} deployments.
The current efforts in modelling datacenters focus on very specialized parts of datacenter management, \ie cooling and heat modelling, network mapping.
Many crucial features, inherent to the \gls{dt} definition are still missing from current \gls{dcdt}s.
Standalone \gls{dcdt} systems fail to offer the holistic capabilities envisioned by the inventors of \gls{dt}s.
The results of the literature survey are in \Cref{tab:dt_features_comparison}, which contains systems we found through a semi-structured literature review process.
We first used structured queries, followed by a mix of snowballing and manual search.
As a result of the findings from \Cref{tab:dt_features_comparison}, we additionally provide a holistic system model that encompasses the features of all the systems from \Cref{tab:dt_features_comparison} (see \Cref{fig:system_model}).
This system model organizes the scattered, standalone systems into a single diagram.
We hope it will prove useful to future researchers navigating the field of datacenter digital twinning.
\item \emph{How to design a reference architecture for a predictive datacenter digital twin using discrete-event simulation?}\\
To answer this research question, we first brainstormed the potential use-cases for a predictive \gls{dcdt}.
The use-cases (see \Cref{sss:use_cases}) are based on the findings of our literature survey.
Based on them, we created a set of functional and non-functional requirements (see \Cref{sss:functional_requirements,sss:non_functional_requirements}) to guide our system design.
Using the \emph{AtLarge Design Process}~\cite{DBLP:conf/icdcs/IosupVTETBFMT19} we created the reference architecture that enables predictive analysis for datacenter operators through digital twinning.
Our system contains 4 major elements: \begin{enumerate*}[label=(\arabic*)]
\item the datacenter (physical twin),
\item the digital thread,
\item the digital twin, and
\item predictive analytics.
\end{enumerate*}
For a detailed discussion of the design, see \Cref{ss:design_of_mysystem}.
\item \emph{How to validate and evaluate a datacenter digital twin architecture in relation to system requirements?}\\
To answer the last research question we crated a prototype.
During the prototype design, we used state-of-the-practice software, such as \code{Confluent Kafka}, \code{Redis} and \code{PostgreSQL} (see \Cref{ss:implementation_overview}).
However, as it turns out, evaluating \gls{dcdt}s is not a trivial task.
Lacking the physical datacenter to experiment with, we came up with a novel digital twin evaluation method.
Our method, relies solely on discrete-event simulation to model the physical datacenter, overcoming the problems of real-world experimentation (\eg sustainability, costliness, reproducibility).
The findings indicate that \gls{my_system} can reliably differentiate between large host failures and insignificant downtime using predictions from \code{OpenDC}, a state-of-the-art datacenter modelling software.
Moreover, we show that \gls{my_system} can be used to incorporate predictive analytics systems and significantly lower the total number of task failures during a workload.
\end{enumerate}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\textwidth]{images/48_years.pdf}
\caption[48 years of microprocessor trend data.]{48 years of microprocessor trend data. Legend: \textcolor{Orange}{$\blacktriangle$ Transistors (thousands)}, \textcolor{Blue}{$\lgblkcircle$ Single Thread Performance (SpecINT $\times 10^3$)}, \textcolor{Green}{$\lgblksquare$ Frequency (MHz)}, \textcolor{Maroon}{$\blacktriangledown$ Typical Power (Watts)}, $\mdlgblkdiamond$ Number of Logical Cores~\cite{DBLP:image/48Microprocessor/Rupp}.}
\label{fig:rupp_48_years_microprocessor_data}
\end{figure}
\section{Future Work}\label{ss:future_work}
The hardware required to run future \gls{ai} models will more heterogeneous and power-hungry than ever before~\cite{DBLP:journals/computer/AthavaleBBMMPS24}.
Due to the end of Dennard's scaling and the fading of Moore's law, we expect datacenter compute to incorporate more sophisticated architectures, (\eg GPUs, NPUs, TPUs) (see \Cref{fig:rupp_48_years_microprocessor_data}).
To tackle datacenter diversification, \gls{dcdt}s are urgently needed.
We envision \gls{dcdt}s as systems that encompass features necessary to model the entire datacenter.
To take advantage of all benefits of \gls{dt}-ing, a \gls{dcdt} must become a holistic, widely-employed tool.
If successful, \gls{dt}s will be indispensable in datacenter management, given the current growth of \gls{ai} and the diversification of compute under way.
To achieve the \gls{nasem} goals of digital twinning~\cite{DBLP:usdoe/report/AP26894}, we suggest several directions future work should take.
\subsection{A New, Strong Principle of \gls{dcdt} Design}\label{sss:future_work_in_analytics}
\begin{enumerate}[label=\textbf{\arabic*.},align=left]
\item \textbf{The Future Goal}\\
\mysystem enables \gls{dcdt}s to incorporate predictive analysis into facility management.
Naturally, we envision upcoming \gls{dcdt}s will use sophisticated prediction techniques (\eg statistical methods, \gls{ml}).
A \gls{dcdt} must posses predictive capabilities, by definition~\cite{DBLP:usdoe/report/AP26894}.
\item \textbf{What Is Missing?}\\
To power the predictions, we envision an \gls{ml}-based inference engine as a necessary component of digital twinning.
The need for \gls{ml} arises naturally in scenarios where large volumes of data, requiring little to no preprocessing meet the demand for estimating future facility behaviour~\cite{Wikipedia:PredictiveModelling,CambridgeUniversityPress:book/Deisenroth}.
However, currently there are no \gls{dcdt} deployments that model the warehouse using an \gls{ml} approach to predict events (see \Cref{tab:dt_features_comparison}).
\item \textbf{The Next Steps}\\
In short, we stipulate \gls{dcdt}s should include \gls{ml} in their \gls{oda} analysis.
The next steps would involve using \mysystem to employ a \gls{ml}-based inference engine.
To achieve this we propose to use battle-tested \gls{ml} algorithms,
\gls{erm}, the \gls{svm}, linear regression, Bayesian liner regression \etc
For future work in failure prediction, we envision an \gls{abc} approach to estimate the real failure distribution within the datacenter.
The clear benefits of this approach stem from the availability and ease of access to large volumes of data.
Moreover, \gls{ml} models are much quicker to train than \gls{cfd}-based models, and have already been employed for other purposes in existing \gls{dcdt}s~\cite{DBLP:conf/noms/ZhangZLZWC22}.
There are few drawbacks to using \gls{ml} in this scenario.
Existing works, envision going a step further, to even employ \gls{ai}-based inference engines~\cite{DBLP:journals/computer/AthavaleBBMMPS24}.
Presently, no other statistical method can approach the accuracy of a good \gls{ml} model.
\end{enumerate}
\subsection{A Crucial Step to Long-Term Success}\label{sss:future_work_in_education}
\begin{enumerate}[label=\textbf{\arabic*.},align=left]
\item \textbf{The Future Goal}\\
In \Cref{ss:future_work} we highlighted the paradigm shift in datacenter design.
Computer Systems is a fast-paced, dynamic field, and we need educated engineers to make correct, well-informed decisions using \gls{dcdt}s.
We stipulate \gls{dcdt}s, specifically using simulation-based experimentation, should be included in higher education and academia (alike discrete-event simulation~\cite{DBLP:conf/ccgrid/MastenbroekAJLB21}).
New, young students should be made aware of the different Computer Systems trade-offs and encouraged to experiment with \gls{dcdt} tools on their own~\cite{DBLP:conf/icdcs/IosupUVAEHTBT18}.
\item \textbf{What Is Missing?}\\
Currently, \mysystem is not an education-ready program.
While a pleasant \gls{ui} exists in the form of a \code{Grafana} dashboard, \mysystem does not facilitate an intuitive \gls{ui} to experiment and explore with the \gls{dcdt} capabilities.
A clear, graphical \gls{ui}, since Douglas Engelbrat first demonstrated it in 1968~\cite{Wikipedia:article/UserInterface}, is an indispensable part of many computer programs.
Presently, it is missing from \mysystem, and from many other \gls{dt} deployments~\cite{DBLP:conf/sensys/LiW0Z0T23, DBLP:conf/noms/ZhangZLZWC22, DBLP:conf/sc/TaheriBPRHDEWPM24, DBLP:conf/sc/BrewerMKWBHSGGW24} (excluding the visualization dashboards).
\item \textbf{The Next Steps}\\
In summary, we propose \gls{dcdt}s be enhanced with a friendly, education-supportive \gls{ui}.
The next steps to achieve this are \gls{dt}-specific.
For example, to include a \gls{ui} with \mysystem we envision exploring the existing Kotlin \gls{ui} libraries (\eg Jetpack Compose~\cite{Wikipedia:article/JetpackCompose}).
We suggested creating either a \gls{gui} or a \gls{tui} for high accessibility.
The benefits of sharing \gls{dcdt}s to higher-education and academia are vast, and there are few to none drawbacks.
Master-level courses on \gls{dt} design already exist (see \url{https://studiegids.vu.nl/en/courses/2026-2027/XMU_0068}).
We believe including \gls{dcdt} experimentation in university courses can only prove beneficial to the Computer Science society.
\end{enumerate}
\subsection{Real-World Survey of Datacenter Digital Twinning}\label{sss:future_work_in_surveying}
\begin{enumerate}[label=\textbf{\arabic*.},align=left]
\item \textbf{The Future Goal}\\
In \Cref{s:background} we surveyed the literature on datacenter digital twinning.
We followed the steps outlined by Kitchenham \etal~\cite{DBLP:journals/infsof/KitchenhamPBBTNL10}, however we did not conduct any real-world interviews.
Real-world practice can be much different from state-of-the-art published in scientific-journals~\cite{DBLP:journals/corr/abs-2103-02060}.
We propose a series of interviews with industry practitioners to gain a fuller insight into the potential use-cases of \gls{dcdt}s.
We stipulate a systematic literature survey, alongside qualitative interviews can offer much benefit to the \gls{dcdt} community.
\item \textbf{What Is Missing?}\\
In this thesis we have conducted a simple, comprehensive literature survey of \gls{dcdt}s.
There already exist systematic literature surveys of generic \gls{dt}s~\cite{DBLP:conf/cirp/TAO2018169}, however what is still missing is a systematic literature survey supported by real-world interviews and careful analysis of the \gls{dcdt} subdomain.
Microsoft already offers digital twinning as a service (see \url{https://azure.microsoft.com/en-us/products/digital-twins/}), and NVIDIA is deploying digital twins as well (see \url{https://www.nvidia.com/en-sg/omniverse/}).
However, there exists little knowledge on state-of-the-practice (\eg datacenters using the 6SigmaDC software, NVIDIA Omniverse, Microsoft Azure Digital Twins).
To develop useful digital twins, engineers must be aware of the practical challenges and any tacit knowledge associated with \gls{dcdt}s.
\item \textbf{The Next Steps}\\
In short, there is no systematic literature survey of \gls{dcdt}s completed alongside real-world interviews.
The next steps to start the literature survey is to plan, conduct and analyze interviews with practitioners from a wide variety of background and nationalities.
For an excellent example of such survey, in the field of datacenter capacity planning, we advise to read ``Capelin: Fast Data-Driven Capacity Planning for Cloud Datacenters'' by Andreadis \etal~\cite{DBLP:journals/corr/abs-2103-02060}.
We believe the scientific research community can greatly benefit from a more holistic view of \gls{dcdt}s.
Most importantly, the survey could provide valuable insight into practical \gls{dcdt} deployments using proprietary tools (\eg Microsoft Azure Digital Twins).
\end{enumerate}
|
