source: doc/theses/thierry_delisle_MMath/text/parallelism.tex @ 4b75ae9

% ######     #    ######     #    #       #       ####### #       ###  #####  #     #
% #     #   # #   #     #   # #   #       #       #       #        #  #     # ##   ##
% #     #  #   #  #     #  #   #  #       #       #       #        #  #       # # # #
% ######  #     # ######  #     # #       #       #####   #        #   #####  #  #  #
% #       ####### #   #   ####### #       #       #       #        #        # #     #
% #       #     # #    #  #     # #       #       #       #        #  #     # #     #
% #       #     # #     # #     # ####### ####### ####### ####### ###  #####  #     #
\chapter{Parallelism}
Historically, computer performance was about processor speeds and instruction counts. However, with heat dissipation being a direct consequence of speed increase, parallelism has become the new source for increased performance~\cite{Sutter05, Sutter05b}. In this decade, it is no longer reasonable to create a high-performance application without caring about parallelism. Indeed, parallelism is an important aspect of performance and more specifically throughput and hardware utilization. The lowest-level approach of parallelism is to use \glspl{kthread} in combination with semantics like \code{fork}, \code{join}, etc. However, since these have significant costs and limitations, \glspl{kthread} are now mostly used as an implementation tool rather than a user oriented one. There are several alternatives to solve these issues that all have strengths and weaknesses. While there are many variations of the presented paradigms, most of these variations do not actually change the guarantees or the semantics, they simply move costs in order to achieve better performance for certain workloads.

\section{Paradigms}
\subsection{User-Level Threads}
A direct improvement on the \gls{kthread} approach is to use \glspl{uthread}. These threads offer most of the same features that the operating system already provides but can be used on a much larger scale. This approach is the most powerful solution as it allows all the features of multithreading, while removing several of the more expensive costs of kernel threads. The downside is that almost none of the low-level threading problems are hidden; users still have to think about data races, deadlocks and synchronization issues. These issues can be somewhat alleviated by a concurrency toolkit with strong guarantees, but the parallelism toolkit offers very little to reduce complexity in itself.

Examples of languages that support \glspl{uthread} are Erlang~\cite{Erlang} and \uC~\cite{uC++book}.

\subsection{Fibers : User-Level Threads Without Preemption} \label{fibers}
A popular variant of \glspl{uthread} is what is often referred to as \glspl{fiber}. However, \glspl{fiber} do not present meaningful semantic differences with \glspl{uthread}. The significant difference between \glspl{uthread} and \glspl{fiber} is the lack of \gls{preemption} in the latter. Advocates of \glspl{fiber} list their high performance and ease of implementation as major strengths, but the performance difference between \glspl{uthread} and \glspl{fiber} is controversial, and the ease of implementation, while true, is a weak argument in the context of language design. Therefore this proposal largely ignores fibers.

An example of a language that uses fibers is Go~\cite{Go}

\subsection{Jobs and Thread Pools}
An approach on the opposite end of the spectrum is to base parallelism on \glspl{pool}. Indeed, \glspl{pool} offer limited flexibility but at the benefit of a simpler user interface. In \gls{pool} based systems, users express parallelism as units of work, called jobs, and a dependency graph (either explicit or implicit) that ties them together. This approach means users need not worry about concurrency but significantly limit the interaction that can occur among jobs. Indeed, any \gls{job} that blocks also block the underlying worker, which effectively means the CPU utilization, and therefore throughput, suffers noticeably. It can be argued that a solution to this problem is to use more workers than available cores. However, unless the number of jobs and the number of workers are comparable, having a significant number of blocked jobs always results in idles cores.

The gold standard of this implementation is Intel's TBB library~\cite{TBB}.

\subsection{Paradigm Performance}
While the choice between the three paradigms listed above may have significant performance implications, it is difficult to pin down the performance implications of choosing a model at the language level. Indeed, in many situations one of these paradigms may show better performance but it all strongly depends on the workload. Having a large amount of mostly independent units of work to execute almost guarantees equivalent performance across paradigms and that the \gls{pool}-based system has the best efficiency thanks to the lower memory overhead (i.e., no thread stack per job). However, interactions among jobs can easily exacerbate contention. User-level threads allow fine-grain context switching, which results in better resource utilization, but a context switch is more expensive and the extra control means users need to tweak more variables to get the desired performance. Finally, if the units of uninterrupted work are large, enough the paradigm choice is largely amortized by the actual work done.

\section{The \protect\CFA\ Kernel : Processors, Clusters and Threads}\label{kernel}
A \gls{cfacluster} is a group of \glspl{kthread} executed in isolation. \Glspl{uthread} are scheduled on the \glspl{kthread} of a given \gls{cfacluster}, allowing organization between \glspl{uthread} and \glspl{kthread}. It is important that \glspl{kthread} belonging to a same \glspl{cfacluster} have homogeneous settings, otherwise migrating a \gls{uthread} from one \gls{kthread} to the other can cause issues. A \gls{cfacluster} also offers a pluggable scheduler that can optimize the workload generated by the \glspl{uthread}.

\Glspl{cfacluster} have not been fully implemented in the context of this thesis. Currently \CFA only supports one \gls{cfacluster}, the initial one.

\subsection{Future Work: Machine Setup}\label{machine}
While this was not done in the context of this thesis, another important aspect of clusters is affinity. While many common desktop and laptop PCs have homogeneous CPUs, other devices often have more heterogeneous setups. For example, a system using \acrshort{numa} configurations may benefit from users being able to tie clusters and/or kernel threads to certain CPU cores. OS support for CPU affinity is now common~\cite{affinityLinux, affinityWindows, affinityFreebsd, affinityNetbsd, affinityMacosx}, which means it is both possible and desirable for \CFA to offer an abstraction mechanism for portable CPU affinity.

\subsection{Paradigms}\label{cfaparadigms}
Given these building blocks, it is possible to reproduce all three of the popular paradigms. Indeed, \glspl{uthread} is the default paradigm in \CFA. However, disabling \gls{preemption} on the \gls{cfacluster} means \glspl{cfathread} effectively become \glspl{fiber}. Since several \glspl{cfacluster} with different scheduling policy can coexist in the same application, this allows \glspl{fiber} and \glspl{uthread} to coexist in the runtime of an application. Finally, it is possible to build executors for thread pools from \glspl{uthread} or \glspl{fiber}, which includes specialized jobs like actors~\cite{Actors}.