PDF 原檔：報告_OCP2026_SiTime_分散式系統時間不確定性_20261015_original.pdf
講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，
### p<n>為頁碼；圖表內容見原檔。
圖片清單（已驗證 2026-10-19）
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_SiTime_分散式系統時間不確定性_20261015_p3.png |
p3 | 真資料圖 | 授時路徑示意：Spine 交換器→ToR 交換器→NIC→CPU，每節點各有時間不確定度界線，端到端為四者相加 |
報告_OCP2026_SiTime_分散式系統時間不確定性_20261015_p5.png |
p5 | 真資料圖 | TTU 公式：±TTU(PE)=TU_Peak（確定性，來自溫度電壓負載老化）＋Q(PE)×TU_RMS（隨機，Allan deviation），PE=N×erfc(Q/√2) |
報告_OCP2026_SiTime_分散式系統時間不確定性_20261015_p7.png |
p7 | 真資料圖 | 振盪器規格表範例資料：隨機分量 TU_RMS 0.13／2.6／82 ns、確定性 TU_Peak 17／181／1856 ns，對應 servo 頻寬 0.1／0.01／0.001 Hz；條件 ±1°C 與 6°C/min（-40 至 105°C） |
原始內容
p1
Kishan Shenoi Sr. Principal System Architect, SiTime Gary Giust Sr. Director System Architecture, SiTime Bounding Time Uncertainty in Distributed Systems TIME APPLIANCES (TAP)
p2
Distributed Systems Bound Time Uncertainty to Prevent Silent Errors Too Conservative Bound is wider than needed Right Sized Bound matches the true uncertainty Too Optimistic Bound is tighter than the true uncertainty Wasted Time Waiting out uncertainty reduces efficiency Timestamp Uncertainty Bound Insignificant Overlap Acceptable Overlap Too Much Overlap Outcome Server A Server B Efficient & Safe Best efficiency without silent data errors Silent Data Errors Probability of error causes wrong ordering with no error reported
p3
End-to-end time uncertainty is a combination of per-node bounds along the path Goal: Quantify & Optimize Time Uncertainty at Network Level
p4
• PTP synchronizes. The data center bounds time uncertainty. • To ensure no silent data errors, the operator must establish a defensible end-to-end bound across oscillator behavior, residual asymmetry, timestamping, servo response, environmental effects, and other worst-case implementation errors. • Start by bounding uncertainty in oscillator datasheet • Today: Difficult to bound uncertainty from datasheet specs • Future: Total Time Uncertainty (TTU) – a single spec for synchronization applications Time is Critical Metric, not Frequency
p5
Total Time Uncertainty (TTU) Must Account for Deployed Fleet of Oscillators ±𝑇𝑇𝑈𝑃𝐸 = 𝑇𝑈𝑃𝑒𝑎𝑘 + 𝑄𝑃𝐸×𝑇𝑈𝑅𝑀𝑆 Deterministic Component Random Component Distribution 𝑃𝐸 2 Noise Source Deployment • Variation in temperature, voltage, load, process • Aging • Allan Deviation • Thermal profile • Airflow profile • Constant temperature • Acceptable annual probability of error, PE •
of oscillators deployed, N
𝑃𝐸= N×erfc Τ 𝑄 2 Solve for Q iteratively
p6
• Transparent clock (switch) – Oscillator TTU contribution adds linearly to network uncertainty • Ordinary clock (NIC) – Oscillator TTU contribution is high-pass filtered by servo loop TTU Calculation – Account for Location of Oscillator in Network - Loop Filter Local Time Local Time Local Clock Frequency correction Timestamp processing Network time stamps Network time Local Oscillator (LO) IEEE 1588 PTP Servo Loop
p7
Oscillator Datasheet Specifies TTU for Data Center Statistics Time Uncertainty (TU) Symbol TU for a given servo bandwidth Condition 0.1 Hz 0.01 Hz 0.001 Hz Random component TURMS 0.13 ns 2.6 ns 82 ns Ambient temperature within ±1 C Deterministic component TUPeak 17 ns 181 ns 1856 ns Any thermal profile up to 6 C/min over -40 C to 105 C Total TU[1] TTUOSC TUPeak + TURMS × Q Q is the “crest factor” that converts a rms to a peak value for probability PE Notes 1. Solve Q iteratively for an acceptable annual probability of error, PE, and deployed fleet of N oscillators: PE = N×erfc(Q/sqrt(2)) 2. TURMS and TUPeak values derived from common oscillator measurements Example Data
p8
Time Uncertainty (TU) from Measurement Analysis Temperature vs Time T (°C) t (sec) T2 T1 Frequency Error vs Temp. F (ppb) T (°C) T1 T2 MEASUREMENT (Sample size = M devices) • Simulate Servo for temperature error and for random error • BW choices: 0.1Hz, 0.01Hz, 0.001Hz • For each choice of BW establish: • Temperature effect: P(k) = |TE|MAX • Random behavior: S(k) = |TE|RMS SIMULATION (for each of k = 1,2,…M devices) STATISTICS Time Error vs Time (Constant Temp.) e (ns) t (sec) Error Signal (temp. fluctuation) Error Signal (random) + • E • Statistically extend results from sample size of M to whole population • For each choice of BW: • Temperature: TUPeak {P(k)} • Random component: TURMS {S(k)}
p9
• Physical layer synchronization • PTP or 1 PPS time transfer • XO → TCXO → OCXO (lower noise oscillators) • EPPS for in-rack and in-server phase distribution Implementations to Reduce TTU
p10
• Challenge – Reduce guardband. Increase efficiency. Preserve correctness. • TTU framework is new, which relies on existing test methods • Contribute • Questions: Standardize TTU in OCP? Typical or maximum values? Thermal profiles? etc. • Offer feedback: ggiust@sitime.com, kshenoi@sitime.com • Get Involved • Project Wiki: https://www.opencompute.org/w/index.php?title=Time_Appliances_Project • Workstream Wiki: https://www.opencompute.org/w/index.php?title=TAP_Oscillators • Mailing list: https://ocp-all.groups.io/g/OCP-TAP Standardize TTU Framework – Components & Equipment
p11
Thank You!