556 lines
No EOL
18 KiB
TeX
556 lines
No EOL
18 KiB
TeX
\documentclass{beamer}
|
|
\usepackage[]{biblatex}
|
|
\usepackage[export]{adjustbox}
|
|
|
|
\title{
|
|
Cache Coherency \& Memory Model in RDMA-Backed Software-Coherent DSM
|
|
}
|
|
\author{Zhengyi Chen}
|
|
\date{\today}
|
|
|
|
\addbibresource{../main.bib}
|
|
|
|
\begin{document}
|
|
% Title Page
|
|
\frame{\titlepage}
|
|
|
|
% Table of Content
|
|
\begin{frame}
|
|
\frametitle{Table of Contents}
|
|
\tableofcontents
|
|
\end{frame}
|
|
|
|
% Part 1: Overview
|
|
% =============================================================================
|
|
\section{1. Overview}
|
|
% Page 1
|
|
\begin{frame}
|
|
\frametitle{1. Overview}
|
|
\begin{itemize}
|
|
\item {
|
|
DSM used to be constrained by NIC bandwidth \& transfer rate (e.g.,
|
|
during the 1990s).
|
|
}
|
|
\item {
|
|
The advent of high(er) transfer rate NICs allows the DSM idea to be
|
|
revived.
|
|
}
|
|
\item {
|
|
Orthogonally, hardware acceleration resources are scarce and highly
|
|
valuable.
|
|
\begin{itemize}
|
|
\item {
|
|
Traditional Scheduling Mechanisms within a Cluster cannot
|
|
dynamically allocate hardware accelerators without high
|
|
overhead.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
Ideally, via high-speed NICs, hardware accelerator could be
|
|
statically allocated such that:
|
|
\begin{itemize}
|
|
\item {
|
|
Every node have access to the hardware accelerator node in a
|
|
time-shared fashion.
|
|
}
|
|
\item {
|
|
Accelerator-attached node can access remote memory much like
|
|
attaching accelerator over, say, PCIe.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Heterogeneous Memory Management}
|
|
\begin{itemize}
|
|
\item {
|
|
\textbf{HMM} facilitates shared address space and transparent data
|
|
migration between CPU and peripherals. Specifically:
|
|
\begin{itemize}
|
|
\item {
|
|
HMM provides interface for duplicating the CPU page table
|
|
with that of the device's, which are transparently
|
|
synchronized.
|
|
}
|
|
\item {
|
|
It also provides corresponding \texttt{struct page}
|
|
representation of device memory pages, which are faulted
|
|
between the CPU and device.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
Theoretically, this should allow for devices in remote nodes to
|
|
perform HMM using the DMA-capable NIC as a ``proxy HMM device''.
|
|
}
|
|
\item {
|
|
Details of implementation of DSM-over-HMM is beyond this thesis's
|
|
scope.
|
|
\begin{itemize}
|
|
\item {
|
|
This thesis focuses on studying and implementing cache
|
|
coherency and later, memory model for the DSM subsystem of
|
|
this wider, ongoing project.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Cache Coherency, and Why It Matters Here}
|
|
\begin{itemize}
|
|
\item {
|
|
Cache-incoherent RDMA (e.g., mlx) performs DMA without
|
|
synchronization with CPU cache.
|
|
}
|
|
\item {
|
|
We cannot assume MMU to magically maintain coherence.
|
|
\begin{itemize}
|
|
\item {
|
|
This seems the case for x86\_64 (cache-coherent DMA), but
|
|
not ARM64.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
At transportation time:
|
|
\begin{itemize}
|
|
\item {
|
|
Send to remote: flushes cache into memory before posting
|
|
send message.
|
|
}
|
|
\item {
|
|
Receive from remote: invalidate cache entry after worked
|
|
recv message.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
Example: Linux kernel tree, \textit{smbdirect} implementation.
|
|
\begin{itemize}
|
|
\item {
|
|
\textit{smbdirect} opportunistically establish SMB over
|
|
RDMA-capable network.
|
|
}
|
|
\item {
|
|
\texttt{smbd\_post\_send} cleans cache entry prior to posting
|
|
send request.
|
|
}
|
|
\item {
|
|
\texttt{recv\_done} invalidates cache entry after exiting
|
|
softirq for recv request (as callback from RDMA driver).
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Consistency Model and Protocol}
|
|
\begin{itemize}
|
|
\item {
|
|
Majority of DSM literatures apply \textbf{release consistency} as
|
|
the system's memory model.
|
|
}
|
|
\item {
|
|
With \textbf{single-writer} protocol, however, the memory model can
|
|
be strengthened with little increase in code complexity.
|
|
\begin{itemize}
|
|
\item {
|
|
\textit{DSPM}\cite{shan2017distributed}, for example,
|
|
achieves a \textit{de-facto} TSO consistency from its
|
|
multi-writer release consistency counterpart -- assuming
|
|
correct memory barriers within each node's CPU, distributed
|
|
writes are never reordered, and distributed reads can
|
|
overtake writes.
|
|
}
|
|
\item {
|
|
Consequently, one can easily achieve sequential consistency
|
|
by designating the entire write-access duration as a critical
|
|
section.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
HMM's ``CPU-or-device'' data migration model also strongly implies
|
|
a single-writer consistency protocol.
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
% Part 2: Design
|
|
% =============================================================================
|
|
\section{2. Design}
|
|
|
|
\begin{frame}
|
|
\frametitle{2. Design}
|
|
\begin{itemize}
|
|
\item {
|
|
Designing a DSM necessitates designing:
|
|
\begin{itemize}
|
|
\item Consistency Model.
|
|
\item Coherence Protocol and State Machine.
|
|
\item Access Control.
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
Care needs to be taken to ensure that the in-kernel implementation
|
|
is:
|
|
\begin{itemize}
|
|
\item Correct,
|
|
\item Performant,
|
|
\item Exploits RDMA's traits.
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Protocol Excerpt: Write-Invalidate}
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\linewidth]{
|
|
w12_slides_resources/Fig-RwlockProtocol 2023-12-06 19_05_06.pdf
|
|
}
|
|
\end{figure}
|
|
The \textit{T}-state indicates a transitionary state for some shared page.
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Consistency Model: TSO}
|
|
\begin{itemize}
|
|
\item {
|
|
Total Store Ordering allows Reads to bypass Stores.
|
|
}
|
|
\item {
|
|
Assuming correct use of node-local synchronization on all nodes,
|
|
applying TSO in a home-based DSM allows for:
|
|
\begin{itemize}
|
|
\item {
|
|
When another node tries to read T-page from access-control
|
|
node: W$\rightarrow$R violation.
|
|
}
|
|
\item {
|
|
When another node tries to read S-page from data-provider
|
|
nodes: W$\rightarrow$R violation (if e.g., the invalidation
|
|
message from access-control node was received afterwards).
|
|
}
|
|
\item {
|
|
Data-provider and access-control nodes work on one request
|
|
at a time: no R$\rightarrow$W violation.
|
|
}
|
|
\item {
|
|
Write-accesses serialized at access-control node: no
|
|
W$\rightarrow$W violation.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Consistency Model: Strengthen to Sequential}
|
|
\begin{itemize}
|
|
\item {
|
|
By corollary, can reverse the previous page's statements to
|
|
strengthen to sequential consistency:
|
|
\begin{itemize}
|
|
\item {
|
|
Disallow T-pages from being serviced until new page content
|
|
is installed: lengthens critical section.
|
|
}
|
|
\item {
|
|
Abolish data-provider nodes: access-control nodes become
|
|
bottleneck.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Coherence Protocol: Possible Features}
|
|
\begin{itemize}
|
|
\item {
|
|
Multi-data-provider Protocol: Instead of having one data-provider,
|
|
have multiple data-provider nodes that are automatically write-back
|
|
to prevent network bottleneck.
|
|
\begin{itemize}
|
|
\item Data provider nodes may be dynamically assigned.
|
|
\item Extra metadata can limit scalability.
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
Auto-share: likewise, write-back pages to non-data-provider nodes to
|
|
take advantage of 1-sided communications.
|
|
}
|
|
\item {
|
|
Request aggregation: aggregate RDMA transfers for optimal transfer
|
|
performance.
|
|
\begin{itemize}
|
|
\item Need to be coherent with program sequence!
|
|
\item Enables write-request merging.
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Stateful Nodes \& Transitions (Provisional)}
|
|
\begin{itemize}
|
|
\item {
|
|
Nodes (e.g., within the cluster) become tightly bound with the
|
|
properties of each shared page(s).
|
|
}
|
|
\end{itemize}
|
|
\begin{figure}
|
|
\centering
|
|
\includegraphics[width=\linewidth]{
|
|
w15_resources/截屏 2024-01-30 19.15.45 2024-01-30 19_16_19.png
|
|
}
|
|
\end{figure}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Stateful Nodes \& Transitions (Provisional) (Cont.)}
|
|
\begin{itemize}
|
|
\item {
|
|
MN (Manager Nodes): Provide access-control and (fallback)
|
|
data-provision.
|
|
}
|
|
\item {
|
|
HN (Home Nodes): Provide data-provision. Can be write-back or
|
|
write-invalidate.
|
|
}
|
|
\item {
|
|
SN (Sharer Nodes): Share data within a reader-only ``epoch''. Can be
|
|
write-back or write-invalidate.
|
|
}
|
|
\item {
|
|
NSN (Non-sharer Nodes): Nodes in network without sharing the
|
|
particular page(s).
|
|
}
|
|
\item {
|
|
CN (Commit Node): Node that acquired the single-writer access to the
|
|
shared page.
|
|
}
|
|
\item {
|
|
Message variants are not finalized:
|
|
\begin{itemize}
|
|
\item {
|
|
Goal: Composable message chains that allow for
|
|
``piggy-backing'' of multiple procedures.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Stateful Nodes: Transition Paths}
|
|
\begin{itemize}
|
|
\item {
|
|
Filled line transitions indicate local requests remote to perform
|
|
state transition.
|
|
}
|
|
\item {
|
|
Dashed line transitions indicate local implicitly transitions prior
|
|
to sending request to remote.
|
|
}
|
|
\item {
|
|
\textit{Non-committal} path concerns about read-only and
|
|
copy-on-write sharing. Sharers cannot make global modification to
|
|
cached local data.
|
|
}
|
|
\item {
|
|
\textit{Invalidation} path is duo with commit operations (due to
|
|
write-invalidation).
|
|
}
|
|
\item {
|
|
\textit{Committal} path concerns about global write sharing. Only
|
|
one writer is allowed to write and commit at one time.
|
|
}
|
|
\item {
|
|
Problem: How exactly to integrate RDMA remote read/write into this?
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
% Part 3: Progress
|
|
% =============================================================================
|
|
\section{3. Progress}
|
|
|
|
\begin{frame}
|
|
\frametitle{3. Progress}
|
|
\begin{itemize}
|
|
\item {
|
|
Goal: in-kernel implementation of software cache-coherency via
|
|
non-coherent RDMA hardware.
|
|
}
|
|
\item {
|
|
Optimistic Goal: in-kernel implementation of memory model in DSM.
|
|
}
|
|
\item {
|
|
Progress: studied and isolated mechanism for data cache
|
|
invalidation/flushing in ARM64, which allows the DSM to run in
|
|
heterogeneous ISA clusters.
|
|
}
|
|
\item {
|
|
Integration with kernel \& main DSM kernel module remains at hand:
|
|
is it absolutely necessary to export new symbols for such an
|
|
important operation?
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{On-demand Coherency in ARM64}
|
|
\begin{itemize}
|
|
\item {
|
|
ARMv8 defines two levels of cache coherence:
|
|
\begin{itemize}
|
|
\item {
|
|
\textit{Point-of-Unification}: Within a core, instruction
|
|
cache, data cache, and TLB all agree in the copy seen for a
|
|
particular address.
|
|
\begin{itemize}
|
|
\item Notably, changing PTE requires PoU.
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
\textit{Point-of-Coherence}: Between all DMA-capable
|
|
peripherals (CPU or otherwise), they all agree in the copy
|
|
seen for a particular address.
|
|
}
|
|
\end{itemize}
|
|
For this thesis's purposes, strive for PoC.
|
|
}
|
|
\item {
|
|
Operations to achieve the latter are encapsulated in the Linux
|
|
kernel as \texttt{(d|i)cache\_(clean|inval)\_poc}.
|
|
\begin{itemize}
|
|
\item Declared under \texttt{arch/arm64/include/asm/cacheflush.h}.
|
|
\item Defined in \texttt{arch/arm64/mm/cache.S}.
|
|
\item {
|
|
Takes virtual address wrt. \textit{current} address space to
|
|
writeback/invalidate cache entries.
|
|
}
|
|
\item {
|
|
Problem: Can only be called in process context (for userspace
|
|
virtual addresses) or in all contexts
|
|
(for kernel virtual addresses)?
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Kernel Patch for On-demand Coherency}
|
|
\begin{itemize}
|
|
\item {
|
|
Problem: These symbols are not exported -- not intended for driver
|
|
use.
|
|
}
|
|
\item {
|
|
Temporary solution: re-export them via patching the kernel.
|
|
\begin{itemize}
|
|
\item Note: Kernel version v6.7.0
|
|
\item {
|
|
Longish-term solution: arrange kernel module code in a way
|
|
that takes advantage of existing driver API
|
|
(e.g., via DMA API, for example \textit{smbdirect}).
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\item {
|
|
Implements wrapper function \texttt{\_\_dcache\_clean\_poc} to
|
|
re-export \texttt{dcache\_clean\_poc} into driver namespace.
|
|
}
|
|
\item {
|
|
Exports symbol into separate header file.
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
\begin{frame}
|
|
\frametitle{Proof-of-Concept Kernel Module}
|
|
\begin{itemize}
|
|
\item {
|
|
Dynamically allocates \texttt{GFP\_USER} pages and remaps to
|
|
userspace on \texttt{mmap}.
|
|
\begin{itemize}
|
|
\item {
|
|
\texttt{GFP\_USER} so (for convenience) pages can be
|
|
directly addressable in kernelspace (via kernel page table).
|
|
}
|
|
\item {
|
|
Pages are lazily allocated and shared between multiple
|
|
processes (i.e., user address spaces).
|
|
}
|
|
\item {
|
|
Exposed as character device \texttt{/dev/my\_shmem}.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\item Around 300+ LoC.
|
|
\item {
|
|
Problem: flawed premise for testing cache writeback!
|
|
\begin{itemize}
|
|
\item {
|
|
Summary: CPU datapath differs from DMA datapath, common cache
|
|
coherency maintenance operations are already performed
|
|
in common file/virtual memory area operation code.
|
|
}
|
|
\item {
|
|
Idea: perform cache write-back on \texttt{vm\_ops->close}.
|
|
}
|
|
\item {
|
|
Reality: virtual memory area already cleaned from cache and
|
|
removed from address space prior to calling
|
|
\texttt{vm\_ops->close}.
|
|
}
|
|
\item {
|
|
Fix: Implement custom \texttt{ioctl}?
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{itemize}
|
|
\end{frame}
|
|
|
|
% Part 4: Future Work
|
|
% =============================================================================
|
|
\section{4. Future Work}
|
|
|
|
\begin{frame}
|
|
\frametitle{4. Future Work}
|
|
\begin{enumerate}
|
|
\item {
|
|
Incorporate cache coherence mechanism into the larger project.
|
|
}
|
|
\item {
|
|
Implement memory model within the larger project. This involves:
|
|
\begin{itemize}
|
|
\item {
|
|
Making adjustment to message type and structure specifications
|
|
for better inter-operation with RDMA.
|
|
}
|
|
\item {
|
|
Implement memory model programmatically.
|
|
}
|
|
\end{itemize}
|
|
}
|
|
\end{enumerate}
|
|
\end{frame}
|
|
|
|
% References
|
|
\begin{frame}
|
|
\frametitle{References}
|
|
\printbibliography
|
|
\end{frame}
|
|
|
|
\end{document} |