Systems Paper Writing

%%%%%%%% ASPLOS 2027 PAPER TEMPLATE %%%%%%%%%%%%%%%%%
%
% ACM International Conference on Architectural Support for
% Programming Languages and Operating Systems
%
% Format: ACM SIGPLAN, <= 12 pages (excluding references), 10pt, two-column
% Uses acmart.cls with sigplan option
%
% Official CFP: https://www.asplos-conference.org/asplos2026/call-for-papers-asplos27/
% ACM Template: https://www.acm.org/publications/proceedings-template
%
% IMPORTANT NOTES:
% - RAPID REVIEW ROUND: Reviewers read ONLY the first 2 pages!
%   --> First 2 pages MUST be self-contained
%   --> Clearly state problem, approach, contribution in first 2 pages
%   --> Do NOT rely on content after page 2 for rapid review
% - Must advance Architecture, PL, and/or OS research
%   --> NOT just using arch/PL/OS to advance another domain
% - Two cycles: April 2026 and September 2026
% - Max 4 submissions per author per cycle
% - Major Revision decision available
% - Double-blind review
%
% RAPID REVIEW TIPS (critical for acceptance):
%   Page 1: Problem motivation + why it matters to arch/PL/OS
%   Page 2: Approach overview + key results preview + contribution list
%   If reviewers cannot determine your contribution to arch/PL/OS
%   from the first 2 pages, your paper WILL be rejected in rapid review.

\documentclass[sigplan,10pt]{acmart}

% Remove copyright/permission footer for submission
\renewcommand\footnotetextcopyrightpermission[1]{}
\settopmatter{printfolios=true}

% Remove ACM reference format for submission
\setcopyright{none}
\renewcommand\acmConference[4]{}
\acmDOI{}
\acmISBN{}

% Recommended packages for architecture/systems papers
\usepackage{booktabs}       % Professional tables
\usepackage{xspace}
\usepackage{subcaption}     % Side-by-side figures
\usepackage{algorithm}      % Algorithm environment
\usepackage{algorithmic}    % Pseudocode formatting
\usepackage{listings}       % Code listings (useful for ISA/compiler examples)
\usepackage[capitalize,noabbrev]{cleveref}  % Smart cross-references

% Code listing style for architecture/compiler papers
\lstset{
  basicstyle=\footnotesize\ttfamily,
  numbers=left,
  numberstyle=\tiny,
  xleftmargin=2em,
  breaklines=true,
  tabsize=2,
  showstringspaces=false,
  frame=single,
  captionpos=b,
  morekeywords={load, store, fence, atomic, sync}  % Add ISA keywords
}

% Custom commands -- replace \system with your anonymized name
\newcommand{\system}{SystemName\xspace}
\newcommand{\eg}{e.g.,\xspace}
\newcommand{\ie}{i.e.,\xspace}
\newcommand{\etal}{\textit{et al.}\xspace}
\newcommand{\para}[1]{\smallskip\noindent\textbf{#1.}}
\newcommand{\parait}[1]{\smallskip\noindent\textit{#1.}}

% Architecture-specific macros
\newcommand{\us}{\,$\mu$s\xspace}
\newcommand{\ns}{\,ns\xspace}
\newcommand{\GHz}{\,GHz\xspace}
\newcommand{\GB}{\,GB\xspace}
\newcommand{\MB}{\,MB\xspace}
\newcommand{\KB}{\,KB\xspace}

\begin{document}

\title{Your Paper Title Here}

% Anonymized for submission
\author{Paper \#XXX}
\affiliation{%
  \institution{Anonymous}
  \country{}}

% Camera-ready (uncomment and fill in):
% \author{Author One}
% \affiliation{%
%   \institution{University/Company}
%   \city{City}
%   \country{Country}}
% \email{email@example.com}
%
% \author{Author Two}
% \affiliation{%
%   \institution{University/Company}
%   \city{City}
%   \country{Country}}
% \email{email@example.com}

\begin{abstract}
% Guidelines for a strong ASPLOS abstract:
% - State what you built/discovered (the contribution)
% - Identify the arch/PL/OS challenge addressed
% - Describe your approach and key insight
% - Quantify improvement with concrete numbers
%
% Keep to 150--200 words. Remember: this is part of the first 2 pages!

We present \system, a [hardware/software/compiler technique] that [capability].
[Problem: why existing arch/PL/OS approaches fall short.]
Our key insight is that [observation about hardware-software interaction].
\system exploits this through [technique], achieving [X]$\times$ speedup
and [Y]\% energy reduction compared to [baseline] on [benchmarks].
\end{abstract}

\maketitle
\pagestyle{plain}

%----------------------------------------------------------------------
% ╔══════════════════════════════════════════════════════════════════════╗
% ║  PAGES 1--2 ARE CRITICAL FOR RAPID REVIEW!                        ║
% ║  Reviewers read ONLY the first 2 pages in the rapid review round. ║
% ║  These must:                                                       ║
% ║  1. Clearly state the problem and why it matters                   ║
% ║  2. Explain how this advances Architecture, PL, or OS              ║
% ║     (NOT just using arch/PL/OS to advance another domain)          ║
% ║  3. Outline your approach and key contributions                    ║
% ║  4. Preview your main results with numbers                         ║
% ╚══════════════════════════════════════════════════════════════════════╝

\section{Introduction}
\label{sec:intro}

% Page 1 should cover: problem motivation + why it matters to arch/PL/OS.
% Page 2 should cover: approach overview + contributions + results preview.

Modern [hardware/software] systems face [challenge] due to
[trend]~\cite{hennessy2019new}. While prior work has addressed
[related problem]~\cite{jouppi2017tpu}, [gap remains].
This paper addresses the [arch/PL/OS] challenge of [specific problem].

\para{Key Insight}
We observe that [insight about hardware-software interaction]. This
observation is supported by our analysis of [N] benchmarks on
[hardware platform] (\cref{sec:background}).

\para{Approach Overview}
\system addresses this through [technique]. Unlike [prior approach],
\system [key distinction], enabling [benefit].

We make the following contributions:
\begin{itemize}
  \item We identify and characterize [problem] through analysis of
        [benchmarks/hardware] (\cref{sec:background}).
  \item We propose \system, a [technique] that [capability]
        (\cref{sec:design}).
  \item We implement \system in [context: compiler/hardware/OS] and
        evaluate on [benchmarks] (\cref{sec:evaluation}).
  \item We demonstrate [X]$\times$ [speedup/efficiency] improvement
        over [state-of-the-art], with only [Y]\% area/power overhead.
\end{itemize}

% End of critical first 2 pages. Content below supports the claims above.
%----------------------------------------------------------------------

\section{Background and Motivation}
\label{sec:background}

\subsection{Hardware/Software Context}

Describe relevant architecture, PL, or OS
background~\cite{lattner2004llvm}.

\subsection{Characterization Study}

% Concrete measurements on real hardware strengthen motivation.
\Cref{fig:motivation} shows [measurement] across [benchmarks] on
[hardware platform].

\begin{figure}[t]
  \centering
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{Performance characterization: breakdown of execution time, \\
    cache miss rates, or energy consumption across benchmarks}
  \vspace{3em}}}
  \caption{Characterization of [metric] across [N] benchmarks on
    [hardware]. On average, [X]\% of [time/energy] is spent on
    [bottleneck], motivating [your approach].}
  \label{fig:motivation}
\end{figure}

\subsection{Opportunity Analysis}

Based on this characterization, we identify [N] key opportunities:

\para{Opportunity 1} [Description with concrete numbers.]

\para{Opportunity 2} [Description.] These opportunities motivate
the design of \system.

%----------------------------------------------------------------------
\section{Design}
\label{sec:design}

\Cref{fig:architecture} shows the overall architecture of \system.

\begin{figure*}[t]
  \centering
  \fbox{\parbox{0.9\textwidth}{\centering\vspace{4em}
    \textit{System architecture: hardware blocks, compiler passes, \\
    or OS mechanisms and their interactions}
  \vspace{4em}}}
  \caption{Architecture of \system. [Describe the key components:
    hardware units, compiler passes, or OS mechanisms.]}
  \label{fig:architecture}
\end{figure*}

\subsection{[Hardware/Compiler/OS Component A]}

Describe the first key component. The core scheduling algorithm is
shown in \cref{alg:scheduling}.

\begin{algorithm}[t]
  \caption{[Algorithm name] in \system}
  \label{alg:scheduling}
  \begin{algorithmic}[1]
    \STATE \textbf{Input:} computation graph $G(V, E)$, resource constraints $R$
    \STATE \textbf{Output:} mapping $M: V \rightarrow R$
    \FOR{each node $v \in \text{TopologicalSort}(V)$}
      \STATE $t_v \leftarrow \max_{(u,v) \in E} (t_u + \text{latency}(u))$
      \STATE $r^* \leftarrow \arg\min_{r \in R} \text{Cost}(v, r, t_v)$
      \STATE $M[v] \leftarrow r^*$
    \ENDFOR
    \STATE \textbf{return} $M$
  \end{algorithmic}
\end{algorithm}

\subsection{[Hardware/Compiler/OS Component B]}

Describe the second component. The performance improvement
from this component can be modeled as:
\begin{equation}
  \label{eq:speedup}
  S = \frac{1}{(1-f) + \frac{f}{p} + \frac{\alpha \cdot f}{B}}
\end{equation}
where $f$ is the parallelizable fraction, $p$ is the number of
processing elements, $B$ is the memory bandwidth, and $\alpha$
is the arithmetic intensity (ops/byte).

% Example: Code transformation (common in ASPLOS PL papers)
\subsection{Example: Code Transformation}

\Cref{fig:transform} shows how \system transforms [code pattern]
to exploit [hardware feature].

\begin{figure}[t]
  \centering
  \begin{minipage}[t]{0.48\columnwidth}
    \centering
    \begin{lstlisting}[title=\textbf{Before},language=C]
for (i = 0; i < N; i++)
  for (j = 0; j < M; j++)
    C[i][j] += A[i][k]
               * B[k][j];
    \end{lstlisting}
  \end{minipage}
  \hfill
  \begin{minipage}[t]{0.48\columnwidth}
    \centering
    \begin{lstlisting}[title=\textbf{After (\system)},language=C]
for (ii = 0; ii < N;
     ii += TILE)
  for (jj = 0; jj < M;
       jj += TILE)
    kernel(A, B, C,
           ii, jj);
    \end{lstlisting}
  \end{minipage}
  \caption{Code transformation example. \system converts [pattern]
    (left) into [optimized pattern] (right), improving [metric]
    by [X]$\times$.}
  \label{fig:transform}
\end{figure}

%----------------------------------------------------------------------
\section{Implementation}
\label{sec:implementation}

We implement \system as follows:
\begin{itemize}
  \item \textbf{[Hardware component]:} [X]K gates in [HDL], synthesized
        at [Y]\GHz using [process node].
  \item \textbf{[Compiler component]:} [X]K lines of [language], integrated
        with [LLVM/GCC/custom compiler].
  \item \textbf{[OS component]:} [X] lines of kernel module in [language].
\end{itemize}

%----------------------------------------------------------------------
\section{Evaluation}
\label{sec:evaluation}

We evaluate \system to answer:
\begin{enumerate}
  \item How does \system compare to state-of-the-art on standard benchmarks?
  \item What is the hardware/software overhead?
  \item How does each component contribute to the improvement?
  \item How sensitive is \system to [key parameters]?
\end{enumerate}

\subsection{Methodology}
\label{sec:eval:method}

\para{Simulation/Hardware}
We evaluate using [simulator/FPGA/real hardware]: [details].

\para{Benchmarks}
We use [SPEC CPU/PARSEC/SPLASH/MLPerf/custom] benchmarks.
\Cref{tab:benchmarks} summarizes the evaluation suite.

\para{Baselines}
We compare against:
(1)~[Baseline A]~\cite{jouppi2017tpu},
(2)~[Baseline B]~\cite{kwon2018maeri}, and
(3)~[Baseline C].

\begin{table}[t]
  \caption{Benchmark suite characteristics.}
  \label{tab:benchmarks}
  \centering
  \begin{small}
  \begin{tabular}{@{}llrr@{}}
    \toprule
    \textbf{Benchmark} & \textbf{Domain} & \textbf{Instructions} &
    \textbf{Working Set} \\
    \midrule
    BenchA  & Image    & 2.1B  & 64\,MB  \\
    BenchB  & NLP      & 4.8B  & 128\,MB \\
    BenchC  & Graph    & 1.3B  & 256\,MB \\
    BenchD  & HPC      & 8.2B  & 512\,MB \\
    BenchE  & Serving  & 0.6B  & 32\,MB  \\
    \bottomrule
  \end{tabular}
  \end{small}
\end{table}

\subsection{Performance Results}
\label{sec:eval:perf}

\Cref{tab:performance} shows the main performance comparison.
\system achieves [X]$\times$ geometric mean speedup over [baseline].

\begin{table}[t]
  \caption{Performance comparison (speedup over baseline).
    Higher is better. Bold indicates best result.}
  \label{tab:performance}
  \centering
  \begin{small}
  \begin{tabular}{@{}lcccc@{}}
    \toprule
    \textbf{Benchmark} & \textbf{Base} & \textbf{Prior A} &
    \textbf{Prior B} & \textbf{\system} \\
    \midrule
    BenchA  & 1.00$\times$ & 1.42$\times$ & 1.55$\times$ & \textbf{2.13}$\times$ \\
    BenchB  & 1.00$\times$ & 1.28$\times$ & 1.39$\times$ & \textbf{1.87}$\times$ \\
    BenchC  & 1.00$\times$ & 1.15$\times$ & 1.22$\times$ & \textbf{1.64}$\times$ \\
    BenchD  & 1.00$\times$ & 1.51$\times$ & 1.68$\times$ & \textbf{2.35}$\times$ \\
    BenchE  & 1.00$\times$ & 1.33$\times$ & 1.41$\times$ & \textbf{1.92}$\times$ \\
    \midrule
    \textit{Geomean} & 1.00$\times$ & 1.33$\times$ & 1.44$\times$ &
    \textbf{1.96}$\times$ \\
    \bottomrule
  \end{tabular}
  \end{small}
\end{table}

\subsection{Area and Power Overhead}
\label{sec:eval:overhead}

\begin{figure}[t]
  \centering
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{Stacked bar chart: area/power breakdown by component}
  \vspace{3em}}}
  \caption{Area and power overhead of \system. The total overhead
    is [X]\% area and [Y]\% power, dominated by [component].}
  \label{fig:overhead}
\end{figure}

\subsection{Ablation Study}
\label{sec:eval:ablation}

\begin{figure}[t]
  \centering
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{Grouped bar chart: performance with components disabled}
  \vspace{3em}}}
  \caption{Ablation study. Removing [Component A] reduces speedup
    from [X]$\times$ to [Y]$\times$, confirming its importance.}
  \label{fig:ablation}
\end{figure}

\subsection{Sensitivity Analysis}
\label{sec:eval:sensitivity}

We vary [key parameter] from [min] to [max] to understand its
impact on performance.

%----------------------------------------------------------------------
\section{Discussion}
\label{sec:discussion}

\para{Generalizability}
[Discuss applicability to other architectures/workloads.]

\para{Limitations}
[Honest discussion of limitations and assumptions.]

%----------------------------------------------------------------------
\section{Related Work}
\label{sec:related}

\para{[Hardware Approaches]}
Prior architecture work~\cite{jouppi2017tpu, kwon2018maeri} addresses
[problem]. \system differs by [distinction].

\para{[Compiler/PL Approaches]}
Compiler techniques~\cite{lattner2004llvm} have targeted [problem].
\system complements these by [distinction].

\para{[OS/Runtime Approaches]}
OS-level approaches~\cite{hennessy2019new} provide [capability].
\system extends this with [technique].

%----------------------------------------------------------------------
\section{Conclusion}
\label{sec:conclusion}

We presented \system, a [technique] that advances [arch/PL/OS area]
by [capability]. \system achieves [X]$\times$ speedup over
state-of-the-art with only [Y]\% overhead, demonstrating the
effectiveness of [key insight].

%----------------------------------------------------------------------
% Acknowledgments (only in camera-ready, remove for submission)
% \begin{acks}
% We thank the anonymous reviewers for their feedback. This work was
% supported by [funding sources].
% \end{acks}

\bibliographystyle{ACM-Reference-Format}
\bibliography{references}

\end{document}

% ASPLOS 2027 Example Bibliography
%
% This file contains example references demonstrating different BibTeX entry
% types commonly used in computer architecture, PL, and OS papers.
% Replace with your actual references.
%
% Entry types demonstrated:
%   inproceedings  -- Conference paper (most common in arch/systems)
%   article        -- Journal article
%   book           -- Book reference
%   phdthesis      -- Doctoral dissertation
%   misc           -- ArXiv preprint or software

%----------------------------------------------------------------------
% Conference papers (inproceedings) -- most common in ASPLOS
%----------------------------------------------------------------------

@inproceedings{jouppi2017tpu,
  author    = {Jouppi, Norman P. and Young, Cliff and Patil, Nishant and
               Patterson, David and Agrawal, Gaurav and Bajwa, Raminder and
               Bates, Sarah and Bhatia, Suresh and Boden, Nan and
               Borber, Al and others},
  title     = {In-Datacenter Performance Analysis of a Tensor Processing Unit},
  booktitle = {Proceedings of the 44th Annual International Symposium on
               Computer Architecture (ISCA)},
  year      = {2017},
  pages     = {1--12},
  address   = {Toronto, ON, Canada},
  publisher = {ACM},
  doi       = {10.1145/3079856.3080246},
}

@inproceedings{kwon2018maeri,
  author    = {Kwon, Hyoukjun and Chatarasi, Parashar and Pellauer, Michael and
               Parashar, Angshuman and Krishna, Tushar and Sarber, Paul},
  title     = {{MAERI}: Enabling Flexible Dataflow Mapping over {DNN}
               Accelerators via Reconfigurable Interconnects},
  booktitle = {Proceedings of the 23rd International Conference on
               Architectural Support for Programming Languages and
               Operating Systems (ASPLOS)},
  year      = {2018},
  pages     = {461--475},
  address   = {Williamsburg, VA},
  publisher = {ACM},
}

@inproceedings{lattner2004llvm,
  author    = {Lattner, Chris and Adve, Vikram},
  title     = {{LLVM}: A Compilation Framework for Lifelong Program Analysis
               and Transformation},
  booktitle = {Proceedings of the International Symposium on Code Generation
               and Optimization (CGO)},
  year      = {2004},
  pages     = {75--86},
  address   = {Palo Alto, CA},
  publisher = {IEEE},
}

@inproceedings{chen2018tvm,
  author    = {Chen, Tianqi and Moreau, Thierry and Jiang, Ziheng and
               Zheng, Lianmin and Yan, Eddie and Sber, Haichen and
               Cowan, Meghan and Wang, Leyuan and Hu, Yuwei and
               Ceze, Luis and Guestrin, Carlos and Krishnamurthy, Arvind},
  title     = {{TVM}: An Automated End-to-End Optimizing Compiler for
               Deep Learning},
  booktitle = {Proceedings of the 13th USENIX Symposium on Operating Systems
               Design and Implementation (OSDI)},
  year      = {2018},
  pages     = {578--594},
  address   = {Carlsbad, CA},
  publisher = {USENIX Association},
}

@inproceedings{barroso2003web,
  author    = {Barroso, Luiz Andr\'{e} and Dean, Jeffrey and H\"{o}lzle, Urs},
  title     = {Web Search for a Planet: The {Google} Cluster Architecture},
  booktitle = {IEEE Micro},
  year      = {2003},
  volume    = {23},
  number    = {2},
  pages     = {22--28},
}

@inproceedings{parashar2019timeloop,
  author    = {Parashar, Angshuman and Raina, Priyanka and Shao, Yakun Sophia
               and Chen, Yu-Hsin and Emer, Joel and others},
  title     = {Timeloop: A Systematic Approach to {DNN} Accelerator Evaluation},
  booktitle = {Proceedings of the IEEE International Symposium on Performance
               Analysis of Systems and Software (ISPASS)},
  year      = {2019},
  pages     = {304--315},
  publisher = {IEEE},
}

%----------------------------------------------------------------------
% Book (book)
%----------------------------------------------------------------------

@book{hennessy2019new,
  author    = {Hennessy, John L. and Patterson, David A.},
  title     = {A New Golden Age for Computer Architecture},
  publisher = {Communications of the ACM},
  year      = {2019},
  volume    = {62},
  number    = {2},
  pages     = {48--60},
  note      = {Turing Award Lecture},
}

%----------------------------------------------------------------------
% ArXiv preprint (misc)
%----------------------------------------------------------------------

@misc{dao2022flashattention,
  author        = {Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and
                   Rudra, Atri and R\'{e}, Christopher},
  title         = {{FlashAttention}: Fast and Memory-Efficient Exact Attention
                   with {IO}-Awareness},
  year          = {2022},
  eprint        = {2205.14135},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
}

%----------------------------------------------------------------------
% PhD thesis (phdthesis)
%----------------------------------------------------------------------

@phdthesis{chen2020dnn,
  author  = {Chen, Yu-Hsin},
  title   = {Efficient Processing of Deep Neural Networks},
  school  = {Massachusetts Institute of Technology},
  year    = {2020},
  address = {Cambridge, MA},
}

%%%%%%%% NSDI 2027 PAPER TEMPLATE %%%%%%%%%%%%%%%%%
%
% The 24th USENIX Symposium on Networked Systems Design and Implementation
% May 11--13, 2027, Providence, RI, USA
%
% Format: <= 12 pages (excluding references), USENIX format
%         Two-column, 10pt on 12pt leading, Times Roman
%
% Official CFP: https://www.usenix.org/conference/nsdi27/call-for-papers
% Template source: https://www.usenix.org/conferences/author-resources/paper-templates
%
% IMPORTANT NOTES:
% - Three tracks: Traditional Research, Frontiers, Operational Systems
% - Indicate track on title page and submission form
% - PRESCREENING PHASE: Reviewers read ONLY the Introduction!
%   --> Introduction must articulate ALL track-specific criteria
% - Two deadlines: Spring (April 2026) and Fall (September 2026)
% - One-shot revision available for rejected papers
%
% TRACK REQUIREMENTS (must be clear from Introduction alone):
%   Research Track:    Novel idea + evaluation evidence
%   Frontiers Track:   Novel NON-INCREMENTAL idea (less evaluation needed)
%   Operational Track: Deployment setting, scale, lessons learned
%
% PRESCREENING CRITERIA (all must be evident in Introduction):
%   1. Subject falls within NSDI scope (networked/distributed systems)
%   2. Exposition understandable by NSDI PC member
%   3. Track-specific criteria met (see above)

\documentclass[letterpaper,twocolumn,10pt]{article}
\usepackage{usenix-2020-09}

% Recommended packages for networking/systems papers
\usepackage[utf8]{inputenc}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{booktabs}       % Professional tables
\usepackage{hyperref}
\usepackage{url}
\usepackage{xspace}
\usepackage{subcaption}     % Side-by-side figures
\usepackage{algorithm}      % Algorithm environment
\usepackage{algorithmic}    % Pseudocode formatting
\usepackage{listings}       % Code listings
\usepackage[capitalize,noabbrev]{cleveref}  % Smart cross-references

% Code listing style
\lstset{
  basicstyle=\footnotesize\ttfamily,
  numbers=left,
  numberstyle=\tiny,
  xleftmargin=2em,
  breaklines=true,
  tabsize=2,
  showstringspaces=false,
  frame=single,
  captionpos=b
}

% Custom commands -- replace \system with your anonymized name
\newcommand{\system}{SystemName\xspace}
\newcommand{\eg}{e.g.,\xspace}
\newcommand{\ie}{i.e.,\xspace}
\newcommand{\etal}{\textit{et al.}\xspace}
\newcommand{\para}[1]{\smallskip\noindent\textbf{#1.}}
\newcommand{\parait}[1]{\smallskip\noindent\textit{#1.}}

% Networking-specific unit macros
\newcommand{\us}{\,$\mu$s\xspace}
\newcommand{\ms}{\,ms\xspace}
\newcommand{\GB}{\,GB\xspace}
\newcommand{\MB}{\,MB\xspace}
\newcommand{\Gbps}{\,Gbps\xspace}
\newcommand{\Tbps}{\,Tbps\xspace}
\newcommand{\pps}{\,pps\xspace}

\begin{document}

% Indicate your track in the title page
% Options: [Research Track] / [Frontiers Track] / [Operational Systems Track]
\title{Your Paper Title Here}

\author{Paper \#XXX}  % Anonymized for submission (double-blind)
% Operational Systems track: may keep real company/system names for context
% Camera-ready:
% \author{
%   {\rm Author One}\\
%   Affiliation One\\
%   \texttt{email@example.com}
%   \and
%   {\rm Author Two}\\
%   Affiliation Two\\
%   \texttt{email@example.com}
% }

\maketitle

%----------------------------------------------------------------------
\begin{abstract}
% Guidelines for a strong NSDI abstract:
% - State the networking/systems problem you solve
% - Explain why existing approaches fail
% - Describe your key insight and approach
% - Summarize evaluation results with concrete numbers
%
% Keep to 150--200 words. Avoid citations in the abstract.

We present \system, a [describe system] for [networked systems problem].
[Problem statement: why existing approaches fall short.]
\system exploits the insight that [key observation] to achieve [capability].
We evaluate \system on [testbed/workloads] and demonstrate [X]$\times$
improvement in [throughput/latency/etc.] compared to [baseline],
while maintaining [other desirable property].
\end{abstract}

%----------------------------------------------------------------------
\section{Introduction}
\label{sec:intro}

% ╔══════════════════════════════════════════════════════════════════╗
% ║  CRITICAL: This section is used for PRESCREENING!              ║
% ║  Reviewers will read ONLY this section to determine:           ║
% ║  1. Subject falls within NSDI scope (networked/distributed)    ║
% ║  2. Exposition understandable by NSDI PC member                ║
% ║  3. Track-specific criteria met (see header comments)          ║
% ║                                                                ║
% ║  If your Introduction doesn't clearly articulate these,        ║
% ║  your paper WILL be rejected in prescreening.                  ║
% ╚══════════════════════════════════════════════════════════════════╝
%
% Recommended structure:
% 1. Problem context in networked/distributed systems (1--2 paragraphs)
% 2. Why existing solutions are insufficient (1 paragraph)
% 3. Key insight and approach overview (1 paragraph)
% 4. Contributions list (bulleted)
% 5. Results highlights with concrete numbers (1 paragraph)

The rapid growth of [networked systems context]~\cite{jain2013b4} has
created new challenges for [problem area]. Existing solutions such as
[prior work]~\cite{alizadeh2010dctcp} are designed for [assumption],
but modern networks require [new capability].

\para{Key Insight}
We observe that [insight about network behavior/workload pattern].
This observation enables \system to [capability].

We make the following contributions:
\begin{itemize}
  \item We characterize [problem] through measurements of [N]
        production [network/cluster] traces (\cref{sec:background}).
  \item We design \system, a [type of system] that leverages
        [technique] to achieve [goal] (\cref{sec:design}).
  \item We implement \system as a [module/protocol/service] with
        [X] lines of [language] (\cref{sec:implementation}).
  \item We evaluate \system on [testbed] with [workloads] and
        show [X]\% improvement in [metric] over state-of-the-art
        (\cref{sec:evaluation}).
\end{itemize}

%----------------------------------------------------------------------
\section{Background and Motivation}
\label{sec:background}

\subsection{Network Architecture Context}

Describe the relevant network architecture or protocol context.
Modern datacenter networks~\cite{singh2015jupiter, greenberg2009vl2}
employ [topology/protocol], which creates [challenge].

\subsection{Measurement Study}

% Concrete measurements from real traces strengthen motivation.
\Cref{fig:motivation} shows [measurement] from [N] production traces.
We identify [N] key findings:

\begin{figure}[t]
  \centering
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{CDF or time-series plot from production trace analysis}
  \vspace{3em}}}
  \caption{[Description.] Analysis of [N] hours of production traffic
    reveals that [finding]: [X]\% of flows account for [Y]\% of bytes.}
  \label{fig:motivation}
\end{figure}

\para{Finding 1}
[First observation from trace analysis.]

\para{Finding 2}
[Second observation.] These findings motivate the design of \system.

%----------------------------------------------------------------------
\section{Design}
\label{sec:design}

\Cref{fig:architecture} presents the architecture of \system.

\begin{figure*}[t]
  \centering
  \fbox{\parbox{0.9\textwidth}{\centering\vspace{4em}
    \textit{System architecture: control plane, data plane, and their interaction}
  \vspace{4em}}}
  \caption{Architecture of \system. The control plane [function] while
    the data plane [function]. [Describe key interactions.]}
  \label{fig:architecture}
\end{figure*}

\subsection{Control Plane}

Describe the control plane design, including how decisions are made
and communicated~\cite{patel2013ananta}.

\subsection{Data Plane}

Describe the data plane design. The forwarding logic is specified
in \cref{alg:forwarding}.

\begin{algorithm}[t]
  \caption{Packet processing in \system}
  \label{alg:forwarding}
  \begin{algorithmic}[1]
    \STATE \textbf{Input:} packet $p$, flow table $F$, policy $\pi$
    \STATE \textbf{Output:} forwarding action $a$
    \STATE $f \leftarrow \text{FlowLookup}(p.\text{header}, F)$
    \IF{$f \neq \text{null}$}
      \STATE $a \leftarrow f.\text{action}$ \COMMENT{cache hit}
    \ELSE
      \STATE $a \leftarrow \pi.\text{Decide}(p)$ \COMMENT{policy lookup}
      \STATE $F.\text{Insert}(p.\text{header}, a)$
    \ENDIF
    \IF{$a.\text{type} = \text{ECMP}$}
      \STATE Select path based on flowlet gap: $\Delta t > \delta$
    \ENDIF
    \STATE \textbf{return} $a$
  \end{algorithmic}
\end{algorithm}

\subsection{Protocol Design}

The bandwidth allocation can be modeled using the max-min fairness
formulation:
\begin{equation}
  \label{eq:fairness}
  \max \min_{i \in \mathcal{F}} \frac{x_i}{w_i}
  \quad \text{s.t.} \quad
  \sum_{i: e \in p_i} x_i \leq c_e, \;\; \forall e \in \mathcal{E}
\end{equation}
where $x_i$ is the rate of flow $i$, $w_i$ is its weight, $p_i$ is
its path, $c_e$ is the capacity of link $e$, and $\mathcal{E}$ is
the set of all links.

\subsection{Handling Failures}

Describe fault tolerance mechanisms. \system handles [failure types]
through [mechanism], achieving [recovery time].

%----------------------------------------------------------------------
\section{Implementation}
\label{sec:implementation}

We implement \system in [X]K lines of [language].

\para{Switch Integration}
[Describe integration with switch hardware/software.]

\para{Host Agent}
[Describe the host-side component.]

\para{Controller}
[Describe the centralized/distributed controller.]

%----------------------------------------------------------------------
\section{Evaluation}
\label{sec:evaluation}

We evaluate \system to answer the following questions:
\begin{enumerate}
  \item Does \system improve [throughput/FCT/latency] over baselines?
  \item How does \system perform under different traffic patterns?
  \item What is the overhead of \system?
  \item How does \system handle failures?
\end{enumerate}

\subsection{Experimental Setup}
\label{sec:eval:setup}

\para{Testbed}
We deploy \system on a [topology] testbed with [N] servers and [M]
switches, connected via [link speed] links.

\para{Traffic Workloads}
We use traffic patterns from [source]~\cite{alizadeh2010dctcp}:
(1)~web search, (2)~data mining, and (3)~cache follower.
\Cref{tab:workloads} summarizes their characteristics.

\para{Baselines}
We compare against:
(1)~ECMP~\cite{hopps2000rfc},
(2)~[Protocol B]~\cite{jain2013b4}, and
(3)~[Protocol C].

\begin{table}[t]
  \caption{Traffic workload characteristics. Flow sizes follow
    the distributions from production datacenter traces.}
  \label{tab:workloads}
  \centering
  \begin{small}
  \begin{tabular}{@{}lrrl@{}}
    \toprule
    \textbf{Workload} & \textbf{Avg Size} & \textbf{Load} &
    \textbf{Distribution} \\
    \midrule
    Web Search   & 1.6\,KB  & 50\%  & Heavy-tailed  \\
    Data Mining   & 7.4\,KB  & 70\%  & Bimodal       \\
    Cache Follow  & 0.4\,KB  & 30\%  & Mostly small  \\
    ML Training   & 128\,MB  & 80\%  & All-to-all    \\
    \bottomrule
  \end{tabular}
  \end{small}
\end{table}

\subsection{Flow Completion Time}
\label{sec:eval:fct}

\Cref{tab:fct} shows flow completion times (FCTs) across workloads.
\system reduces the average FCT by [X]\% and the 99th-percentile
tail FCT by [Y]\% compared to [best baseline].

\begin{table}[t]
  \caption{Flow completion time comparison (normalized to ECMP).
    Lower is better. Bold indicates best result.}
  \label{tab:fct}
  \centering
  \begin{small}
  \begin{tabular}{@{}lcccc@{}}
    \toprule
    & \multicolumn{2}{c}{\textbf{Web Search}} &
      \multicolumn{2}{c}{\textbf{Data Mining}} \\
    \cmidrule(lr){2-3} \cmidrule(lr){4-5}
    \textbf{System} & \textbf{Avg} & \textbf{p99} &
    \textbf{Avg} & \textbf{p99} \\
    \midrule
    ECMP             & 1.00          & 1.00          & 1.00 & 1.00 \\
    Baseline B       & 0.85          & 0.78          & 0.88 & 0.82 \\
    Baseline C       & 0.82          & 0.71          & 0.84 & 0.75 \\
    \textbf{\system} & \textbf{0.68} & \textbf{0.52} & \textbf{0.72} & \textbf{0.58} \\
    \bottomrule
  \end{tabular}
  \end{small}
\end{table}

\subsection{Throughput Under Load}
\label{sec:eval:throughput}

\Cref{fig:throughput} shows aggregate throughput as network load
increases from 10\% to 90\%.

\begin{figure}[t]
  \centering
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{Line chart: throughput vs.\ network load (10\%--90\%)}
  \vspace{3em}}}
  \caption{Aggregate throughput vs.\ network load. \system maintains
    [X]\% of bisection bandwidth at 80\% load, compared to
    [Y]\% for [baseline].}
  \label{fig:throughput}
\end{figure}

\subsection{Failure Recovery}
\label{sec:eval:failure}

We evaluate recovery time by failing [N] links during peak load.
\system recovers within [X]\ms, compared to [Y]\ms for [baseline].

\begin{figure}[t]
  \centering
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{Time-series: throughput drop and recovery after link failure}
  \vspace{3em}}}
  \caption{Failure recovery. \system detects the failure within [X]\us
    and reroutes affected flows within [Y]\ms.}
  \label{fig:failure}
\end{figure}

%----------------------------------------------------------------------
\section{Discussion}
\label{sec:discussion}

\para{Deployment Considerations}
[Discuss practical deployment aspects.]

\para{Limitations}
[Honestly discuss limitations.]

%----------------------------------------------------------------------
\section{Related Work}
\label{sec:related}

% Organize by theme, clearly distinguish your work.

\para{Datacenter Transport Protocols}
DCTCP~\cite{alizadeh2010dctcp} and its successors address [aspect].
\system differs by [distinction].

\para{Traffic Engineering}
B4~\cite{jain2013b4} and Jupiter~\cite{singh2015jupiter} optimize
[aspect]. \system complements these by [distinction].

\para{Load Balancing}
[Other approaches]~\cite{hopps2000rfc, patel2013ananta} provide
[capability]. \system extends this with [technique].

%----------------------------------------------------------------------
\section{Conclusion}
\label{sec:conclusion}

We presented \system, a [type of system] that [key capability].
By exploiting [insight], \system achieves [X]$\times$ improvement
in [metric] over state-of-the-art. Our evaluation on [testbed]
with [workloads] demonstrates [key results].

%----------------------------------------------------------------------
{\footnotesize \bibliographystyle{acm}
\bibliography{references}}

\end{document}

% NSDI 2027 Example Bibliography
%
% This file contains example references demonstrating different BibTeX entry
% types commonly used in networking and distributed systems papers.
% Replace with your actual references.
%
% Entry types demonstrated:
%   inproceedings  -- Conference paper (most common)
%   article        -- Journal article
%   techreport     -- RFC / Technical report
%   phdthesis      -- Doctoral dissertation
%   misc           -- ArXiv preprint, website, or software

%----------------------------------------------------------------------
% Conference papers (inproceedings)
%----------------------------------------------------------------------

@inproceedings{alizadeh2010dctcp,
  author    = {Alizadeh, Mohammad and Greenberg, Albert and Maltz, David A.
               and Padhye, Jitendra and Patel, Parveen and Prabhakar, Balaji
               and Sengupta, Sudipta and Sridharan, Murari},
  title     = {Data Center {TCP} ({DCTCP})},
  booktitle = {Proceedings of the ACM SIGCOMM 2010 Conference},
  year      = {2010},
  pages     = {63--74},
  address   = {New Delhi, India},
  publisher = {ACM},
  doi       = {10.1145/1851182.1851192},
}

@inproceedings{greenberg2009vl2,
  author    = {Greenberg, Albert and Hamilton, James R. and Jain, Navendu and
               Kandula, Srikanth and Kim, Changhoon and Lahiri, Parantap and
               Maltz, David A. and Patel, Parveen and Sengupta, Sudipta},
  title     = {{VL2}: A Scalable and Flexible Data Center Network},
  booktitle = {Proceedings of the ACM SIGCOMM 2009 Conference},
  year      = {2009},
  pages     = {51--62},
  address   = {Barcelona, Spain},
  publisher = {ACM},
}

@inproceedings{jain2013b4,
  author    = {Jain, Sushant and Kumar, Alok and Mandal, Subhasree and
               Ong, Joon and Poutievski, Leon and Singh, Arjun and
               Venkata, Subbaiah and Wanderer, Jim and Zhou, Junlan and
               Zhu, Min and Zolla, Jon and H\"{o}lzle, Urs and Stuart, Stephen
               and Vahdat, Amin},
  title     = {{B4}: Experience with a Globally-Deployed Software Defined {WAN}},
  booktitle = {Proceedings of the ACM SIGCOMM 2013 Conference},
  year      = {2013},
  pages     = {3--14},
  address   = {Hong Kong, China},
  publisher = {ACM},
}

@inproceedings{patel2013ananta,
  author    = {Patel, Parveen and Bansal, Deepak and Yuan, Lihua and
               Murthy, Ashwin and Greenberg, Albert and Maltz, David A.
               and Kern, Randy and Kumar, Hemant and Zikos, Marios and
               Wu, Hongyu and Kim, Changhoon and Karri, Naveen},
  title     = {Ananta: Cloud Scale Load Balancing},
  booktitle = {Proceedings of the ACM SIGCOMM 2013 Conference},
  year      = {2013},
  pages     = {207--218},
  address   = {Hong Kong, China},
  publisher = {ACM},
}

@inproceedings{singh2015jupiter,
  author    = {Singh, Arjun and Ong, Joon and Agarwal, Amit and Anderson, Glen
               and Armistead, Ashby and Bannon, Roy and Boving, Seb and
               Desai, Gaurav and Felderman, Bob and Germano, Paulie and others},
  title     = {Jupiter Rising: A Decade of {Clos} Topologies and Centralized
               Control in {Google}'s Datacenter Network},
  booktitle = {Proceedings of the ACM SIGCOMM 2015 Conference},
  year      = {2015},
  pages     = {183--197},
  address   = {London, UK},
  publisher = {ACM},
}

@inproceedings{handley2017quic,
  author    = {Langley, Adam and Riddoch, Alistair and Wilk, Alyssa and
               Vicente, Antonio and Krasic, Charles and Zhang, Dan and
               Yang, Fan and Kouranov, Fedor and Swett, Ian and Iyengar, Janardhan
               and others},
  title     = {The {QUIC} Transport Protocol: Design and Internet-Scale Deployment},
  booktitle = {Proceedings of the ACM SIGCOMM 2017 Conference},
  year      = {2017},
  pages     = {183--196},
  address   = {Los Angeles, CA},
  publisher = {ACM},
}

%----------------------------------------------------------------------
% Journal article (article)
%----------------------------------------------------------------------

@article{floyd1993random,
  author    = {Floyd, Sally and Jacobson, Van},
  title     = {Random Early Detection Gateways for Congestion Avoidance},
  journal   = {IEEE/ACM Transactions on Networking},
  volume    = {1},
  number    = {4},
  pages     = {397--413},
  year      = {1993},
  doi       = {10.1109/90.251892},
  publisher = {IEEE},
}

%----------------------------------------------------------------------
% RFC / Technical report (techreport)
%----------------------------------------------------------------------

@techreport{hopps2000rfc,
  author      = {Hopps, Christian E.},
  title       = {Analysis of an Equal-Cost Multi-Path Algorithm},
  institution = {Internet Engineering Task Force},
  year        = {2000},
  type        = {RFC},
  number      = {2992},
  note        = {\url{https://www.rfc-editor.org/rfc/rfc2992}},
}

%----------------------------------------------------------------------
% ArXiv preprint (misc)
%----------------------------------------------------------------------

@misc{netllm2024,
  author        = {Wu, Duo and Wang, Xianda and Qiao, Yaqi and
                   Wang, Zhi and Jiang, Junchen and Cui, Shuguang and
                   Wang, Fangxin},
  title         = {{NetLLM}: Adapting Large Language Models for Networking},
  year          = {2024},
  eprint        = {2402.02338},
  archivePrefix = {arXiv},
  primaryClass  = {cs.NI},
}

%----------------------------------------------------------------------
% PhD thesis (phdthesis)
%----------------------------------------------------------------------

@phdthesis{alizadeh2013thesis,
  author  = {Alizadeh, Mohammad},
  title   = {Large Scale Transport for Data Centers},
  school  = {Stanford University},
  year    = {2013},
  address = {Stanford, CA},
}

% USENIX style file for papers
% usenix-2020-09.sty
%
% This is the official USENIX style for conferences including OSDI, NSDI, ATC, etc.
% Source: https://www.usenix.org/conferences/author-resources/paper-templates
%
% NOTE: This is a simplified version for template purposes.
% For the latest official version, download from:
% https://www.usenix.org/conferences/author-resources/paper-templates

\NeedsTeXFormat{LaTeX2e}
\ProvidesPackage{usenix-2020-09}[2020/09/01 USENIX Style]

% Required packages
\RequirePackage{mathptmx}      % Times Roman font
\RequirePackage[scaled=0.92]{helvet} % Helvetica for sans-serif
\RequirePackage{courier}       % Courier for monospace
\RequirePackage{graphicx}
\RequirePackage{url}

% Page layout: 7" x 9" text block on 8.5" x 11" paper
\setlength{\textheight}{9.0in}
\setlength{\textwidth}{7.0in}
\setlength{\columnsep}{0.33in}
\setlength{\topmargin}{0.0in}
\setlength{\headheight}{0.0in}
\setlength{\headsep}{0.0in}
\setlength{\oddsidemargin}{-0.25in}
\setlength{\evensidemargin}{-0.25in}
\setlength{\parindent}{1em}
\setlength{\parskip}{0pt}

% Title formatting
\renewcommand{\@maketitle}{%
  \newpage
  \null
  \vskip 2em%
  \begin{center}%
    \let \footnote \thanks
    {\LARGE \@title \par}%
    \vskip 1.5em%
    {\large
      \lineskip .5em%
      \begin{tabular}[t]{c}%
        \@author
      \end{tabular}\par}%
    \vskip 1em%
    {\large \@date}%
  \end{center}%
  \par
  \vskip 1.5em}

% Section formatting
\renewcommand{\section}{\@startsection{section}{1}{\z@}%
  {-3.5ex \@plus -1ex \@minus -.2ex}%
  {2.3ex \@plus.2ex}%
  {\normalfont\large\bfseries}}

\renewcommand{\subsection}{\@startsection{subsection}{2}{\z@}%
  {-3.25ex\@plus -1ex \@minus -.2ex}%
  {1.5ex \@plus .2ex}%
  {\normalfont\normalsize\bfseries}}

\renewcommand{\subsubsection}{\@startsection{subsubsection}{3}{\z@}%
  {-3.25ex\@plus -1ex \@minus -.2ex}%
  {1.5ex \@plus .2ex}%
  {\normalfont\normalsize\bfseries}}

% Footnote formatting
\renewcommand{\thefootnote}{\fnsymbol{footnote}}

% Abstract formatting
\renewenvironment{abstract}%
  {\begin{quote}\small\textbf{Abstract: }}%
  {\end{quote}}

% Float parameters
\renewcommand{\topfraction}{0.9}
\renewcommand{\bottomfraction}{0.8}
\renewcommand{\textfraction}{0.1}
\renewcommand{\floatpagefraction}{0.8}

\endinput

%%%%%%%% OSDI 2026 PAPER TEMPLATE %%%%%%%%%%%%%%%%%
%
% The 20th USENIX Symposium on Operating Systems Design and Implementation
% July 13--15, 2026, Seattle, WA, USA
%
% Format: <= 12 pages (excluding references), 8.5"x11", 10pt on 12pt leading,
%         two-column, Times Roman, 7"x9" text block
% Camera-ready: <= 14 pages (2 extra pages allowed)
%
% Official CFP: https://www.usenix.org/conference/osdi26/call-for-papers
% Template source: https://www.usenix.org/conferences/author-resources/paper-templates
%
% IMPORTANT NOTES:
% - OSDI 2026 has two tracks: Research and Operational Systems
% - For Operational Systems track, title must end with "(Operational Systems)"
% - Max 8 submissions per author
% - Papers should be the right length (not padded to 12 pages)
% - Papers <= 6 pages are unlikely to receive full consideration
% - Use anonymized project/system name (different from arXiv/talks)
%
% WHAT OSDI REVIEWERS LOOK FOR:
% 1. Significant problem motivation
% 2. Interesting and compelling solution
% 3. Practicality and benefits demonstrated
% 4. Clear contribution articulation
% 5. Advances beyond previous work

\documentclass[letterpaper,twocolumn,10pt]{article}
\usepackage{usenix-2020-09}

% Recommended packages for systems papers
\usepackage[utf8]{inputenc}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{booktabs}       % Professional tables
\usepackage{hyperref}
\usepackage{url}
\usepackage{xspace}
\usepackage{subcaption}     % Side-by-side figures
\usepackage{algorithm}      % Algorithm environment
\usepackage{algorithmic}    % Pseudocode formatting
\usepackage{listings}       % Code listings
\usepackage[capitalize,noabbrev]{cleveref}  % Smart cross-references

% Code listing style for systems papers
\lstset{
  basicstyle=\footnotesize\ttfamily,
  numbers=left,
  numberstyle=\tiny,
  xleftmargin=2em,
  breaklines=true,
  tabsize=2,
  showstringspaces=false,
  frame=single,
  captionpos=b
}

% Custom commands -- replace \system with your anonymized name
\newcommand{\system}{SystemName\xspace}
\newcommand{\eg}{e.g.,\xspace}
\newcommand{\ie}{i.e.,\xspace}
\newcommand{\etal}{\textit{et al.}\xspace}
\newcommand{\para}[1]{\smallskip\noindent\textbf{#1.}}
\newcommand{\parait}[1]{\smallskip\noindent\textit{#1.}}

% Convenience macros for units (common in systems papers)
\newcommand{\us}{\,$\mu$s\xspace}
\newcommand{\ms}{\,ms\xspace}
\newcommand{\GB}{\,GB\xspace}
\newcommand{\MB}{\,MB\xspace}
\newcommand{\Gbps}{\,Gbps\xspace}

\begin{document}

% For submission: use anonymized title and Paper #XXX as author
% For Operational Systems track: add "(Operational Systems)" to title
\title{Your Paper Title Here}
% \title{Your Paper Title Here (Operational Systems)}  % Operational Systems track

\author{Paper \#XXX}  % Anonymized for submission
% Camera-ready:
% \author{
%   {\rm Author One}\\
%   Affiliation One\\
%   \texttt{email@example.com}
%   \and
%   {\rm Author Two}\\
%   Affiliation Two\\
%   \texttt{email@example.com}
% }

\maketitle

%----------------------------------------------------------------------
\begin{abstract}
% Guidelines for a strong OSDI abstract:
% - State what you achieved (the contribution)
% - Why this is hard and important (the problem)
% - How you do it (the approach)
% - What evidence you have (evaluation highlights)
% - Your most remarkable result (the hook)
%
% Keep to 150--200 words. Avoid citations in the abstract.

We present \system, a [describe system] that [key capability].
[Problem statement: why existing approaches fall short.]
\system addresses this through [key technique/insight].
We evaluate \system on [workloads/benchmarks] and show that it achieves
[X]\% improvement in [metric] over [baseline], while reducing [other metric]
by [Y]$\times$.
\end{abstract}

%----------------------------------------------------------------------
\section{Introduction}
\label{sec:intro}

% Structure your introduction as follows:
% 1. Problem context and motivation (1--2 paragraphs)
% 2. Why existing solutions are insufficient (1 paragraph)
% 3. Key insight / approach overview (1 paragraph)
% 4. Contributions (bulleted list)
% 5. Results highlights (1 paragraph)
%
% OSDI reviewers look for: significant problem + compelling solution +
% demonstrated practicality + clear contributions + advances beyond prior work.

Modern systems face the challenge of [describe problem]~\cite{dean2004mapreduce}.
As workloads grow in scale and complexity, existing approaches such as
[prior work]~\cite{abadi2016tensorflow} struggle to [limitation].

\para{Key Insight}
Our key observation is that [insight]. This enables \system to [capability]
without [drawback of prior approaches].

We make the following contributions:
\begin{itemize}
  \item We identify [problem/opportunity] and characterize its impact
        on [workloads] (\cref{sec:background}).
  \item We design \system, which introduces [technique] to address
        [challenge] (\cref{sec:design}).
  \item We implement \system in [X] lines of [language] and integrate
        it with [existing system] (\cref{sec:implementation}).
  \item We evaluate \system on [benchmarks] and demonstrate [X]$\times$
        improvement over [baseline] (\cref{sec:evaluation}).
\end{itemize}

%----------------------------------------------------------------------
\section{Background and Motivation}
\label{sec:background}

% Provide context the reader needs to understand your contribution.
% Include a motivating example or measurement study.

\subsection{Problem Context}

Describe the system context and relevant background.
Prior work~\cite{moritz2018ray, zaharia2012spark} has explored [related area],
but [gap remains].

\subsection{Motivating Example}

% Use concrete numbers from real workloads to motivate the problem.
\Cref{fig:motivation} shows [measurement] across [workloads].
We observe that [finding], which motivates our approach.

\begin{figure}[t]
  \centering
  % Replace with your actual figure
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{Motivating measurement or characterization study}
  \vspace{3em}}}
  \caption{[Description of motivating measurement.] We observe that
    [key finding] across [N] production workloads, motivating the need
    for [your approach].}
  \label{fig:motivation}
\end{figure}

%----------------------------------------------------------------------
\section{Design}
\label{sec:design}

% Present your system design top-down:
% 1. Architecture overview (with figure)
% 2. Key components / mechanisms
% 3. How they interact

\Cref{fig:architecture} shows the overall architecture of \system.
The system consists of [N] key components: [list].

\begin{figure*}[t]
  \centering
  % Replace with your actual architecture diagram
  \fbox{\parbox{0.9\textwidth}{\centering\vspace{4em}
    \textit{System architecture diagram showing key components and data flow}
  \vspace{4em}}}
  \caption{Architecture of \system. [Component A] handles [function],
    while [Component B] manages [function]. Arrows indicate [data/control flow].}
  \label{fig:architecture}
\end{figure*}

\subsection{Component A: [Name]}

Describe the first key component. A formal specification of the core
algorithm can be found in \cref{alg:core}.

\begin{algorithm}[t]
  \caption{Core algorithm of \system}
  \label{alg:core}
  \begin{algorithmic}[1]
    \STATE \textbf{Input:} workload $W$, resources $R$
    \STATE \textbf{Output:} scheduling plan $P$
    \STATE Initialize plan $P \leftarrow \emptyset$
    \FOR{each task $t_i \in W$}
      \STATE Estimate resource demand $d_i \leftarrow \text{Predict}(t_i)$
      \IF{$\text{Available}(R) \geq d_i$}
        \STATE $P \leftarrow P \cup \{(t_i, \text{Allocate}(R, d_i))\}$
      \ELSE
        \STATE Enqueue $t_i$ for deferred scheduling
      \ENDIF
    \ENDFOR
    \STATE \textbf{return} $P$
  \end{algorithmic}
\end{algorithm}

\subsection{Component B: [Name]}

Describe the second key component. The expected throughput can be
modeled as:
\begin{equation}
  \label{eq:throughput}
  T = \frac{N \cdot B}{L + \frac{B}{C}}
\end{equation}
where $N$ is the number of parallel workers, $B$ is the batch size,
$L$ is the network latency, and $C$ is the per-worker compute rate.

\subsection{Handling Edge Cases}

Discuss how the design handles failures, stragglers, or other
edge cases important in production systems.

%----------------------------------------------------------------------
\section{Implementation}
\label{sec:implementation}

% Describe implementation details that matter for reproducibility.
% Include system size, language, key libraries, and integration points.

We implement \system in approximately [X]K lines of [language].
Key implementation details include:

\para{Threading Model}
[Describe the threading/concurrency model.]

\para{Integration}
\system integrates with [existing system] by [method of integration].
We modify [N] lines of the original codebase.

%----------------------------------------------------------------------
\section{Evaluation}
\label{sec:evaluation}

% Structure your evaluation to answer specific questions:
% - Q1: How does \system compare to state-of-the-art? (end-to-end)
% - Q2: What is the contribution of each component? (ablation)
% - Q3: How does \system scale? (scalability)
% - Q4: What is the overhead? (cost analysis)

We evaluate \system to answer the following questions:
\begin{enumerate}
  \item How does \system compare to state-of-the-art systems?
  \item What is the contribution of each design component?
  \item How does \system scale with increasing workload?
  \item What overhead does \system introduce?
\end{enumerate}

\subsection{Experimental Setup}
\label{sec:eval:setup}

\para{Testbed}
We run experiments on a cluster of [N] machines, each with
[CPU model], [X]\GB RAM, and [GPU model if applicable],
connected via [network].

\para{Workloads}
We use [N] workloads from [source]: [list workloads].
\Cref{tab:workloads} summarizes their characteristics.

\para{Baselines}
We compare against [N] baselines:
(1)~[Baseline A]~\cite{verma2015borg},
(2)~[Baseline B]~\cite{ongaro2014raft}, and
(3)~[Baseline C].

\begin{table}[t]
  \caption{Workload characteristics used in evaluation.
    [Describe what the columns represent.]}
  \label{tab:workloads}
  \centering
  \begin{small}
  \begin{tabular}{@{}lrrrl@{}}
    \toprule
    \textbf{Workload} & \textbf{Tasks} & \textbf{Data (GB)} &
    \textbf{Duration} & \textbf{Type} \\
    \midrule
    WorkloadA  & 1,024  & 128  & 2.4\,h  & Batch     \\
    WorkloadB  & 512    & 64   & 1.1\,h  & Streaming \\
    WorkloadC  & 4,096  & 512  & 8.7\,h  & ML Train  \\
    WorkloadD  & 256    & 32   & 0.5\,h  & Serving   \\
    \bottomrule
  \end{tabular}
  \end{small}
\end{table}

\subsection{End-to-End Performance}
\label{sec:eval:e2e}

\Cref{tab:e2e} shows the end-to-end performance comparison.
\system achieves [X]\% higher throughput and [Y]\% lower latency
compared to [best baseline].

\begin{table}[t]
  \caption{End-to-end performance comparison. Bold indicates best result.
    \system achieves the highest throughput and lowest p99 latency
    across all workloads.}
  \label{tab:e2e}
  \centering
  \begin{small}
  \begin{tabular}{@{}lccc@{}}
    \toprule
    \textbf{System} & \textbf{Throughput} & \textbf{p50 Latency} &
    \textbf{p99 Latency} \\
                     & \textbf{(Kops/s)}  & \textbf{(ms)}        &
    \textbf{(ms)}        \\
    \midrule
    Baseline A       & 125.3               & 4.2                  & 18.7 \\
    Baseline B       & 142.1               & 3.8                  & 15.2 \\
    Baseline C       & 98.6                & 5.1                  & 22.4 \\
    \textbf{\system} & \textbf{187.4}      & \textbf{2.9}         & \textbf{9.8} \\
    \bottomrule
  \end{tabular}
  \end{small}
\end{table}

\subsection{Ablation Study}
\label{sec:eval:ablation}

To understand the contribution of each component, we evaluate variants
of \system with individual components disabled.
\Cref{fig:ablation} shows the results.

\begin{figure}[t]
  \centering
  % Replace with your ablation study figure
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{Bar chart: \system vs.\ variants with components disabled}
  \vspace{3em}}}
  \caption{Ablation study results. Each bar represents \system with
    one component removed. Component A contributes [X]\% and
    Component B contributes [Y]\% of the total improvement.}
  \label{fig:ablation}
\end{figure}

\subsection{Scalability}
\label{sec:eval:scale}

We evaluate how \system scales from [N] to [M] nodes.
As shown in \cref{fig:scalability}, \system achieves near-linear
scaling up to [K] nodes.

\begin{figure}[t]
  \centering
  \fbox{\parbox{0.9\columnwidth}{\centering\vspace{3em}
    \textit{Line chart: throughput vs.\ number of nodes for each system}
  \vspace{3em}}}
  \caption{Scalability comparison. \system achieves [X]\% of ideal
    linear scaling at [K] nodes, compared to [Y]\% for [baseline].}
  \label{fig:scalability}
\end{figure}

%----------------------------------------------------------------------
\section{Discussion}
\label{sec:discussion}

% Discuss limitations, lessons learned, and generalizability.
% OSDI reviewers appreciate honest discussion of limitations.

\para{Limitations}
[Discuss known limitations of your system.]

\para{Lessons Learned}
[Share insights from building and deploying the system.]

%----------------------------------------------------------------------
\section{Related Work}
\label{sec:related}

% Organize by theme, NOT paper-by-paper.
% Clearly distinguish your work from each category.

\para{[Category A] Systems}
Prior work on [category]~\cite{dean2004mapreduce, abadi2016tensorflow}
focuses on [aspect]. \system differs by [distinction].

\para{[Category B] Approaches}
[Other approaches]~\cite{lamport1978time, verma2015borg} address [problem]
through [method]. In contrast, \system [distinction].

%----------------------------------------------------------------------
\section{Conclusion}
\label{sec:conclusion}

We presented \system, a [description] that [key capability].
Through [technique], \system achieves [improvement] over
state-of-the-art systems. Our evaluation on [workloads] demonstrates
[key results]. [Optional: future work direction.]

%----------------------------------------------------------------------
% Bibliography
% USENIX uses the acm bibliography style
{\footnotesize \bibliographystyle{acm}
\bibliography{references}}

%----------------------------------------------------------------------
% Optional: Appendix (after bibliography for USENIX)
% \appendix
% \section{Additional Evaluation Results}
% Include supplementary material here.

\end{document}

% USENIX style file for papers
% usenix-2020-09.sty
%
% This is the official USENIX style for conferences including OSDI, NSDI, ATC, etc.
% Source: https://www.usenix.org/conferences/author-resources/paper-templates
%
% NOTE: This is a simplified version for template purposes.
% For the latest official version, download from:
% https://www.usenix.org/conferences/author-resources/paper-templates

\NeedsTeXFormat{LaTeX2e}
\ProvidesPackage{usenix-2020-09}[2020/09/01 USENIX Style]

% Required packages
\RequirePackage{mathptmx}      % Times Roman font
\RequirePackage[scaled=0.92]{helvet} % Helvetica for sans-serif
\RequirePackage{courier}       % Courier for monospace
\RequirePackage{graphicx}
\RequirePackage{url}

% Page layout: 7" x 9" text block on 8.5" x 11" paper
\setlength{\textheight}{9.0in}
\setlength{\textwidth}{7.0in}
\setlength{\columnsep}{0.33in}
\setlength{\topmargin}{0.0in}
\setlength{\headheight}{0.0in}
\setlength{\headsep}{0.0in}
\setlength{\oddsidemargin}{-0.25in}
\setlength{\evensidemargin}{-0.25in}
\setlength{\parindent}{1em}
\setlength{\parskip}{0pt}

% Title formatting
\renewcommand{\@maketitle}{%
  \newpage
  \null
  \vskip 2em%
  \begin{center}%
    \let \footnote \thanks
    {\LARGE \@title \par}%
    \vskip 1.5em%
    {\large
      \lineskip .5em%
      \begin{tabular}[t]{c}%
        \@author
      \end{tabular}\par}%
    \vskip 1em%
    {\large \@date}%
  \end{center}%
  \par
  \vskip 1.5em}

% Section formatting
\renewcommand{\section}{\@startsection{section}{1}{\z@}%
  {-3.5ex \@plus -1ex \@minus -.2ex}%
  {2.3ex \@plus.2ex}%
  {\normalfont\large\bfseries}}

\renewcommand{\subsection}{\@startsection{subsection}{2}{\z@}%
  {-3.25ex\@plus -1ex \@minus -.2ex}%
  {1.5ex \@plus .2ex}%
  {\normalfont\normalsize\bfseries}}

\renewcommand{\subsubsection}{\@startsection{subsubsection}{3}{\z@}%
  {-3.25ex\@plus -1ex \@minus -.2ex}%
  {1.5ex \@plus .2ex}%
  {\normalfont\normalsize\bfseries}}

% Footnote formatting
\renewcommand{\thefootnote}{\fnsymbol{footnote}}

% Abstract formatting
\renewenvironment{abstract}%
  {\begin{quote}\small\textbf{Abstract: }}%
  {\end{quote}}

% Float parameters
\renewcommand{\topfraction}{0.9}
\renewcommand{\bottomfraction}{0.8}
\renewcommand{\textfraction}{0.1}
\renewcommand{\floatpagefraction}{0.8}

\endinput