Where everything went每樣東西搬到了哪裡

For readers of the first edition: where each old section now lives, where each concept from the review's candidate list is defined, which statements were corrected against their sources, and which digests could not be checked.給讀過第一版的讀者：每個舊的節現在在哪裡、審查清單中的每個概念在哪裡定義、哪些敘述對照原文做了修正，以及哪些研究導讀無法核對來源。

First-edition sections第一版的各節

The overview's fourteen sections and the eleven deep dives were merged into one sequence. A section that appears once in the table may have been split; the link goes to where its opening now sits.

總覽的十四節與十一個深入章節合併成一條順序。表中的一節可能被拆開了；連結指向它的開頭現在所在的位置。

Overview總覽

First-edition section第一版的節	Now現在
Why AI training needs a special network為什麼 AI 訓練需要專用網路	Section 1.2第 1.2 節
The parts: GPU, server, NIC, switch, link基本元件：GPU、伺服器、NIC、交換器、鏈路	Section 2.1第 2.1 節
How the model is split across GPUs模型如何切分到多顆 GPU	Section 3.6第 3.6 節
Collective communication集體通訊	Section 4.2第 4.2 節
Topology: how the switches are wired拓樸：交換器怎麼接線	Section 6.1第 6.1 節
Routing and load balancing路由與負載平衡	Section 8.1第 8.1 節
Transport: RDMA over Ethernet傳輸層：RDMA over Ethernet	Section 7.1第 7.1 節
Front-end and back-end networks前端網路與後端網路	Section 9.1第 9.1 節
Failures and checkpoints故障與檢查點	Chapter 10第 10 章
Sharing the cluster, and simulating it多租戶共用叢集，以及模擬	Chapter 11第 11 章
Three production designs, side by side三家量產設計並列比較	Chapter 13第 13 章
Guided readings, one paper at a time論文導讀：一次一篇	Optional reading選讀
Reading list and courses閱讀清單與課程	Reading list閱讀清單
Glossary名詞表	Glossary名詞表

Why training needs its own network為什麼訓練需要自己的網路

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
The training step as a timeline把訓練步畫成時間線	Section 1.1第 1.1 節
How much arithmetic: the 6ND rule多少算術：6ND 法則	Section 1.2第 1.2 節
How many bytes per step每一步要搬多少位元組	Section 5.3第 5.3 節
The communication fraction grows with the cluster通訊比例隨叢集變大而增加	Section 5.4第 5.4 節
Synchronization is a barrier, and the barrier waits for the slowest同步是一道柵欄，柵欄等的是最慢的那一個	Section 1.4第 1.4 節
What the traffic looks like on the wire線上的流量長什麼樣	Section 5.6第 5.6 節
What "utilization" means, and what one point is worth「使用率」是什麼意思，一個百分點值多少	Section 1.3第 1.3 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	What clusters and jobs actually carry: Acme and Hermod叢集與任務實際承載了什麼：Acme 與 Hermod
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

The hardware: GPU, server, NIC, switch and optics硬體：GPU、伺服器、NIC、交換器與光學

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，又漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
The bandwidth ladder as a principle頻寬階梯作為原則	Section 2.2第 2.2 節
The server: eight GPUs, two CPUs, nine NICs伺服器：八顆 GPU、兩顆 CPU、九張 NIC	Section 2.1第 2.1 節
The NIC: a transport protocol in siliconNIC：做在矽裡的傳輸協定	Section 2.4第 2.4 節
The switch: one chip, 512 lanes, a port budget交換器：一顆晶片、512 條 lane、一份埠預算	Section 6.4第 6.4 節
Optics and cables: reach, cost, power and failure光學與線纜：距離、成本、功耗與故障	Section 2.6第 2.6 節
The latency ladder and the bandwidth-delay product延遲階梯與頻寬延遲乘積	Section 2.7第 2.7 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	Measuring inside the server: FabricPerf, InfiniteHBD, PingPoint量測伺服器內部：FabricPerf、InfiniteHBD、PingPoint
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Splitting the model: five parallelisms and the traffic they make切分模型：五種平行化與它們產生的流量

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
Why split at all: memory per parameter為何非切不可：每個參數的記憶體	Section 3.1第 3.1 節
Data parallelism and ZeRO資料平行與 ZeRO	Section 3.3第 3.3 節
Tensor parallelism: collectives inside every layer張量平行：每一層內部的集體操作	Section 3.4第 3.4 節
Pipeline parallelism: stages, micro-batches and bubbles管線平行：階段、micro-batch 與氣泡	Section 3.5第 3.5 節
Context parallelism and expert parallelism上下文平行與專家平行	Section 3.6第 3.6 節
Putting them together: a bandwidth order合在一起：一個頻寬順序	Section 3.7第 3.7 節
How a state-of-the-art MoE model is trained目前最先進的 MoE 模型怎麼訓練	Section 3.8第 3.8 節
What changes if如果改變了什麼	Section 3.9第 3.9 節
Recent research at SIGCOMM, NSDI, ISCA and EuroSysSIGCOMM、NSDI、ISCA 與 EuroSys 的近期研究	Parallelism research: when ranks are not equal平行化研究：當各 rank 並不相同
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Collectives and NCCL: algorithms, cost models and the wire集體通訊與 NCCL：演算法、成本模型與線路上的行為

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，又漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
The vocabulary, with sizes詞彙，附上大小	Section 3.2第 3.2 節
The α–β cost modelα–β 成本模型	Section 4.3第 4.3 節
The ring, one step at a time環，一步一步看	Section 4.2第 4.2 節
Hierarchical and topology-aware collectives階層式與拓樸感知的集體操作	Section 4.4第 4.4 節
Inside NCCLNCCL 內部	Section 4.5第 4.5 節
Bytes in flight and latency在途位元組與延遲	Section 4.6第 4.6 節
Overlap: hiding the collective重疊：把集體操作藏起來	Section 4.7第 4.7 節
Measuring: algbw, busbw and efficiency量測：algbw、busbw 與效率	Section 4.8第 4.8 節
Research frontier: what recent papers on collectives work on研究前沿：近期集體通訊論文在做什麼	Collectives research: solvers, new algorithms, the library, MoE and compression集體通訊研究：求解器、新演算法、函式庫、MoE 與壓縮
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Topology: Clos arithmetic, rails, planes and tori拓樸：Clos 的算術、rail、plane 與環面

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，又漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
Clos arithmetic: from one chip to a fabricClos 的算術：從一顆晶片到整個網路	Section 6.1第 6.1 節
Oversubscription and bisection: where to be cheap超額認購與對分：該在哪裡省	Section 6.2第 6.2 節
Rail-optimized, plain, and rail-onlyRail-optimized、一般接線與 rail-only	Section 6.3第 6.3 節
Redundancy by wiring用接線做備援	Section 6.5第 6.5 節
Beyond the Clos: torus, optical circuit switches, dragonflyClos 之外：環面、光路交換器與 dragonfly	Section 6.9第 6.9 節
Physical constraints that size the fabric決定網路規模的物理限制	Section 6.10第 6.10 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	Topology research: wire smarter, or re-wire with optics拓樸研究：更聰明地接線，或用光學重新接線
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Routing and load balancing: why hashing fails and what replaces it路由與負載平衡：為何雜湊失效，以及取而代之的方法

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
What a switch does with a packet交換器如何處理一個封包	Section 8.1第 8.1 節
The arithmetic of collisions碰撞的算術	Section 8.2第 8.2 節
Meta's five stages, and how each one failedMeta 的五個階段，以及每一個如何失敗	Section 8.3第 8.3 節
Alibaba's structural answer: remove the choice阿里的結構性答案：拿掉選擇	Section 8.4第 8.4 節
Per-packet load balancing: spraying and its price逐封包負載平衡：噴灑與它的代價	Section 8.5第 8.5 節
Entropy you can control, and entropy you cannot你能控制的熵，與你不能控制的熵	Section 8.6第 8.6 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	Load-balancing research: in the switch, in the NIC, and watching負載平衡研究：在交換器、在 NIC，以及監看
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Transport: lossless Ethernet, congestion control, and life after PFC傳輸層：無損 Ethernet、壅塞控制，以及 PFC 之後

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
What RoCEv2 puts on the wireRoCEv2 在線上放了什麼	Section 7.1第 7.1 節
Loss recovery decides everything丟包恢復決定一切	Section 7.2第 7.2 節
PFC in detail: headroom, thresholds, pathologiesPFC 細談：headroom、門檻與病症	Section 7.3第 7.3 節
DCQCN in detail: mark, notify, reactDCQCN 細談：標記、通知、反應	Section 7.4第 7.4 節
Why the AI fabrics diverge from the recipe為何 AI 網路偏離這份配方	Section 7.5第 7.5 節
After PFC: fix the NIC, cap the bytes in flightPFC 之後：修 NIC、限制在途位元組	Section 7.6第 7.6 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	Transport research: congestion control for collectives, and reliability without PFC傳輸層研究：為集體操作設計的壅塞控制，以及不靠 PFC 的可靠性
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

The front-end network: data, checkpoints and control前端網路：資料、檢查點與控制

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
Two networks, four kinds of traffic兩個網路，四種流量	Section 9.1第 9.1 節
Dataset traffic: smaller than you think資料集流量：比你想的小	Section 9.2第 9.2 節
Checkpoint traffic: the burst that sizes everything檢查點流量：決定一切的突發	Section 9.3第 9.3 節
Storage systems in one page一頁看完儲存系統	Section 9.4第 9.4 節
Control and management控制與管理	Section 9.5第 9.5 節
Sizing the front end決定前端網路的規模	Section 9.6第 9.6 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	Checkpoint and front-end research: ByteCheckpoint, Checkmate, DistDPU檢查點與前端網路研究：ByteCheckpoint、Checkmate、DistDPU
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Failures, checkpoints and stragglers: the arithmetic of staying up故障、檢查點與 straggler：維持運轉的算術

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，又漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
What fails, and how often什麼會壞、多常壞	Section 10.1第 10.1 節
Why one failure stops everything為什麼一個故障會讓一切停下	Section 10.2第 10.2 節
Checkpoint arithmetic檢查點的算術	Section 10.3第 10.3 節
Automated recovery loops自動化的恢復迴圈	Section 10.4第 10.4 節
Designing so that failures hurt less設計成讓故障傷害更小	Section 10.5第 10.5 節
Stragglers: the cost of the slowest rankStraggler：最慢 rank 的成本	Section 10.6第 10.6 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	Failure research: is it the network, and where?故障研究：是不是網路，問題在哪裡？
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Sharing the cluster: isolation, placement and simulation共享叢集：隔離、配置與模擬

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
Why clusters are shared, and what sharing breaks為何叢集被共享，以及共享會弄壞什麼	Section 11.1第 11.1 節
Placement: the strongest network lever配置：最強的網路槓桿	Section 11.2第 11.2 節
Job shapes, fragmentation and spare capacity任務形狀、碎片化與備用容量	Section 11.3第 11.3 節
Simulation and modeling: step time before the hardware模擬與建模：在硬體之前算出步時間	Section 12.2第 12.2 節
Operating a shared fabric營運共享的網路	Section 11.4第 11.4 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	Sharing and simulation research: schedule, be fair, place, simulate共用與模擬研究：排程、公平、擺放、模擬
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Four production designs compared, and how to choose四種生產設計的比較，以及如何選擇

First-edition section第一版的節	Now現在
What the overview said, and what it left out總覽說了什麼，又漏了什麼	Replaced by the “Before you start” box that opens each chapter改為每章開頭的「開始之前」
The comparison matrix比較矩陣	Section 13.1第 13.1 節
Axis (a): any-to-any, rail-only, or torus軸 (a)：任意對任意、rail-only，還是環面	Section 13.2第 13.2 節
Axis (b): deep buffers or shallow single chips軸 (b)：深緩衝還是淺的單晶片	Section 13.3第 13.3 節
Axis (c): fix load balance in software, in wiring, or in circuits軸 (c)：用軟體、接線，還是電路來修負載平衡	Section 13.4第 13.4 節
Axis (d): where redundancy lives軸 (d)：備援放在哪裡	Section 13.5第 13.5 節
Where the four designs agree四種設計一致的地方	Section 13.6第 13.6 節
A decision procedure決策程序	Section 13.7第 13.7 節
What changed after these papers這些論文之後改變了什麼	Section 13.8第 13.8 節
Recent research at SIGCOMM and NSDISIGCOMM 與 NSDI 的近期研究	Design research: Astral and Pegasus設計研究：Astral 與 Pegasus
Key numbers關鍵數字	Dropped; each number stays in the text where it is derived已移除；每個數字留在推導它的內文中
Terms used on this page本頁用到的術語	Replaced by “New terms in this chapter” and the glossary改為「本章新名詞」與名詞表
Related guided readings相關論文導讀	Listed at the end of each chapter列在每章結尾
Check your understanding檢驗你的理解	The quiz at the end of each chapter每章結尾的測驗

Concept identifiers C01–C62概念編號 C01–C62

The review used these identifiers for its candidate concepts. They are kept here only, as a map to the term that now carries each concept and its single home.

審查以這些編號標記候選概念。它們只保留在這裡，作為對照：每個概念現在由哪個名詞承載、它唯一的定義處在哪裡。

ID	Concept概念	Defined at (chapter)定義處（章）
C01	Rail-only	Rail-onlyRail-only 6
C02	rail-optimized	Rail-optimized wiringRail 最佳化接線 6
C03	rail	RailRail 6
C04	Clos/fat tree	Clos network, fat treeClos 網路、胖樹（fat tree） 6
C05	scale-up/scale-out	Scale-up, scale-outScale-up、scale-out 2
C06	HBD/NVLink	High-bandwidth domain (HBD)高頻寬域（HBD） 2 · NVLink, NVSwitchNVLink、NVSwitch 2
C07	oversubscription	Oversubscription超額認購 2
C08	Dual-ToR	Dual-ToR雙 ToR（dual-ToR） 6
C09	dual-plane	Dual-plane雙平面（dual-plane） 6
C10	torus	Torus環面（torus） 6
C11	optical circuit switching	Optical circuit switch (OCS)光路交換器（OCS） 6
C12	network cost	Section 6.10第 6.10 節
C13	data parallelism	Data parallelism (DP)資料平行（DP） 3
C14	tensor parallelism	Tensor parallelism (TP)張量平行（TP） 3
C15	pipeline parallelism	Pipeline parallelism (PP), stage管線平行（PP）、階段 3
C16	expert parallelism	Expert parallelism (EP)專家平行（EP） 3
C17	context/sequence parallelism	Context parallelism (CP)上下文平行（CP） 3 · Sequence parallelism (SP)序列平行（SP） 3
C18	MoE all-to-all	All-to-allAll-to-all 3 · Dispatch, combine分派（dispatch）、合併（combine） 3
C19	traffic burstiness	Burstiness, duty cycle突發性、工作週期 5
C20	communication-compute overlap	Overlap, exposed communication重疊、暴露的通訊 4
C21	model FLOPs utilization	MFUMFU 1
C22	all-reduce	All-reduceAll-reduce 3
C23	reduce-scatter/all-gather	Reduce-scatterReduce-scatter 3 · All-gatherAll-gather 3
C24	all-to-all	All-to-allAll-to-all 3
C25	ring	Ring all-reduce環狀 all-reduce 4
C26	tree	Tree, recursive doubling and halving樹、遞迴加倍與減半 4
C27	Rabenseifner	Rabenseifner's algorithmRabenseifner 演算法 4
C28	alpha-beta cost	α–β cost modelα–β 成本模型 4
C29	NCCL	NCCLNCCL 4
C30	collective synthesis	Collectives research: solvers, new algorithms, the library, MoE and compression集體通訊研究：求解器、新演算法、函式庫、MoE 與壓縮
C31	compression	Collectives research: solvers, new algorithms, the library, MoE and compression集體通訊研究：求解器、新演算法、函式庫、MoE 與壓縮
C32	ECMP	ECMPECMP 8
C33	hash polarization	Hash polarization雜湊極化 8
C34	flowlet	Flowlet, adaptive routingFlowlet、自適應路由 8
C35	packet spraying	Packet spraying, reordering封包噴灑、亂序 8
C36	RDMA/RoCE	RDMARDMA 2 · RoCEv2RoCEv2 7
C37	queue pair	Queue pair (QP)佇列對（QP） 2
C38	PFC	PFCPFC 7
C39	ECN	ECN, CNPECN、CNP 7
C40	DCQCN	DCQCNDCQCN 7
C41	reordering	Packet spraying, reordering封包噴灑、亂序 8
C42	Go-back-N	Go-back-N, selective retransmissionGo-back-N、選擇性重傳 2
C43	selective retransmission	Go-back-N, selective retransmissionGo-back-N、選擇性重傳 2
C44	front-end/back-end	Front-end and back-end networks前端網路與後端網路 9
C45	checkpointing	Checkpoint檢查點（checkpoint） 9
C46	failure rate/MTBF	MTBFMTBF（平均故障間隔） 10
C47	straggler	Straggler落後者（straggler） 1
C48	fault diagnosis	Telemetry遙測（telemetry） 9 · Section 10.4第 10.4 節
C49	recovery	Section 10.4第 10.4 節
C50	checkpoint interval	Young–Daly intervalYoung–Daly 間隔 10
C51	multi-tenancy	Multi-tenancy, tenant多租戶、租戶 11
C52	placement	Placement, scheduler放置、排程器 11
C53	isolation	Isolation, contention隔離、爭用 11
C54	congestion	Congestion, congestion control壅塞、壅塞控制 2
C55	simulation	Simulation, packet-level simulation模擬、封包層級模擬 12
C56	packet-level simulation	Simulation, packet-level simulation模擬、封包層級模擬 12
C57	trace replay	Section 12.3第 12.3 節
C58	fair scheduling	Sharing and simulation research: schedule, be fair, place, simulate共用與模擬研究：排程、公平、擺放、模擬
C59	collective resilience	Failure research: is it the network, and where?故障研究：是不是網路，問題在哪裡？
C60	benchmarking	algbw, busbwalgbw、busbw 4 · Section 12.4第 12.4 節
C61	bounded outstanding bytes / receiver admission	Bytes in flight在途位元組 2 · Clear-to-send (CTS), receiver-driven admission允許傳送（CTS）、接收端准入 4 · BDP-FCBDP-FC 7
C62	shared DPU control plane across RNICs	DPUDPU 2 · Checkpoint and front-end research: ByteCheckpoint, Checkmate, DistDPU檢查點與前端網路研究：ByteCheckpoint、Checkmate、DistDPU

Corrections against the sources對照原文的修正

Page頁面	What the text now says內文現在的說法
Rail-only	Figures 2–4 and the 8.2% / 11.2% overheads are computed by the authors' analytical model, not measured; Table II is not monotonic.Figure 2–4 與 8.2% / 11.2% 的開銷是作者解析模型的計算值，不是量測；Table II 不是單調的。
Alibaba HPN	Hosts send ARP to both ToRs, which turn it into /32 routes advertised with BGP; the host runs no BGP. The 6.25% drop came from failing one NIC–ToR link while training LLaMa-7B on 256 GPUs. Table 3 does not say over what period its traffic is counted.主機把 ARP 送給兩台 ToR，由 ToR 轉成 /32 路由並以 BGP 宣告；主機不跑 BGP。6.25% 的下降來自在 256 顆 GPU 上訓練 LLaMa-7B 時弄斷一條 NIC–ToR 鏈路。Table 3 沒有說明流量是多長時間內的量。
Llama 3	Table 5's printed percentages and its counts disagree: GPU-related causes are 58.7% as printed and 64.0% by count; the first row is 35.3%, printed as 30.1%.Table 5 的印刷百分比與次數不一致：GPU 相關依印刷值為 58.7%，依次數為 64.0%；第一列是 35.3%，印成 30.1%。
Crux	Theorem 1's limit is 1, under the proof's conditions (repeated pairing, a fixed finite set of jobs, one bottleneck link).Theorem 1 的極限是 1，且只在證明的前提下成立（重複的配對、固定且有限的任務集合、一條瓶頸鏈路）。
DCP	The control queue is designed not to drop, but the paper's own simulation lost 0.16% of header-only packets at a 255-to-1 incast without congestion control.控制佇列的設計目標是不丟包，但論文自己的模擬在沒有壅塞控制、255 對 1 的 incast 下仍丟了 0.16% 的純標頭封包。
R-Pingmesh	Checked against the full paper (14 pages) in October 2026: 12 text changes in the digest and 2 on chapter or frontier pages.2026 年 10 月對照論文全文（14 頁）核對：導讀內文修改 12 處，各章或研究前沿頁修改 2 處。
Hermod	Checked against the full paper (3 pages) in October 2026: 8 text changes in the digest and 4 on chapter or frontier pages.2026 年 10 月對照論文全文（3 頁）核對：導讀內文修改 8 處，各章或研究前沿頁修改 4 處。
From ATOP to ZCube	Checked against the full paper (21 pages) in October 2026: 19 text changes in the digest and 4 on chapter or frontier pages.2026 年 10 月對照論文全文（21 頁）核對：導讀內文修改 19 處，各章或研究前沿頁修改 4 處。
Balancing and Beyond	Checked against the full paper (14 pages) in October 2026: 38 text changes in the digest and 2 on chapter or frontier pages.2026 年 10 月對照論文全文（14 頁）核對：導讀內文修改 38 處，各章或研究前沿頁修改 2 處。
Balanced Sparse Tree	Checked against the full paper (20 pages) in October 2026: 19 text changes in the digest.2026 年 10 月對照論文全文（20 頁）核對：導讀內文修改 19 處。
DistDPU	Checked against the full paper (16 pages) in October 2026: 10 text changes in the digest and 2 on chapter or frontier pages.2026 年 10 月對照論文全文（16 頁）核對：導讀內文修改 10 處，各章或研究前沿頁修改 2 處。
FabricPerf	Checked against the full paper (15 pages) in October 2026: 18 text changes in the digest and 4 on chapter or frontier pages.2026 年 10 月對照論文全文（15 頁）核對：導讀內文修改 18 處，各章或研究前沿頁修改 4 處。
HyNA	Checked against the full paper (15 pages) in October 2026: 10 text changes in the digest.2026 年 10 月對照論文全文（15 頁）核對：導讀內文修改 10 處。
LEVELLER	Checked against the full paper (20 pages) in October 2026: 8 text changes in the digest and 2 on chapter or frontier pages.2026 年 10 月對照論文全文（20 頁）核對：導讀內文修改 8 處，各章或研究前沿頁修改 2 處。
Odin	Checked against the full paper (18 pages) in October 2026: 21 text changes in the digest and 4 on chapter or frontier pages.2026 年 10 月對照論文全文（18 頁）核對：導讀內文修改 21 處，各章或研究前沿頁修改 4 處。
OptCCL	Checked against the full paper (13 pages) in October 2026: 23 text changes in the digest and 2 on chapter or frontier pages.2026 年 10 月對照論文全文（13 頁）核對：導讀內文修改 23 處，各章或研究前沿頁修改 2 處。
Pegasus	Checked against the full paper (15 pages) in October 2026: 21 text changes in the digest and 2 on chapter or frontier pages.2026 年 10 月對照論文全文（15 頁）核對：導讀內文修改 21 處，各章或研究前沿頁修改 2 處。
PReCCL	Checked against the full paper (19 pages) in October 2026: 20 text changes in the digest and 2 on chapter or frontier pages.2026 年 10 月對照論文全文（19 頁）核對：導讀內文修改 20 處，各章或研究前沿頁修改 2 處。
Theseus	Checked against the full paper (20 pages) in October 2026: 11 text changes in the digest and 2 on chapter or frontier pages.2026 年 10 月對照論文全文（20 頁）核對：導讀內文修改 11 處，各章或研究前沿頁修改 2 處。