Kodan: Addressing the Computational Bottleneck in SpaceKodan：解決太空中的運算瓶頸
On-board inference creates a deadline for each camera frame. Kodan uses context-specialized models, tiling and selective processing to meet that deadline on embedded hardware. Its gain measures the high-value fraction of a saturated downlink.機上推論為每幀影像增加處理期限。Kodan 使用情境特化模型、影像切塊與選擇性處理，讓嵌入式硬體趕上期限。其增益衡量飽和下行資料流中的高價值比例。
Kodan starts where orbital edge computing left off. Before launch, Kodan groups a representative dataset into geographic contexts and trains specialized models. It selects tiles per frame to balance precision and execution time. It elides processing for tiles whose context is mostly high-value.
These are downlinked, or mostly low-value. These are discarded, with a selection logic that picks the combination for each tile [83, §3] . The execution times were measured on ground hardware and the orbit was simulated, so all results come from ground measurements and simulation.
Kodan 從軌道邊緣運算停下的地方開始。Kodan 的答案是在發射前依目標硬體改造應用：把代表性資料集聚類成地理情境並訓練特化模型，選擇每幀切成幾個圖塊以精確度換取時間，並對情境以高價值為主的圖塊直接下傳、以低價值為主的圖塊直接丟棄而略過處理，再由選擇邏輯為每個圖塊挑出組合[83, §3]。執行時間是在地面硬體上量測、軌道則是模擬的，所以論文中所有數字皆來自地面量測與模擬。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | Once a satellite filters its own imagery before downlinking, the limit moves from the radio to the on-board processor. Landsat 8 observes a new frame every 22 s, while the real cloud filter requires 98 s. Direct filtering increases high-value downlink by only 9%. An ideal filter would give a threefold increase. [83, §2.1.3, Fig. 5] |
| Platforms and evaluation | Seven cloud-segmentation applications run on three platforms. They use a 15 W Jetson AGX Orin, a 140 W Core i7-7800X and a 180 W GTX 1070 Ti. The cote simulator models Landsat 8 orbit, camera and ground infrastructure. Kodan raises saturated-downlink data-value density by 89–97% over the bent pipe. Full ground-track coverage needs up to 12× fewer satellites than prior OEC pipelines. [83, §4, §5.1–5.2, Figs. 8, 11] |
| 影像處理瓶頸 | 衛星一旦在下傳前先過濾自己的影像，限制就從無線電移到機上處理器，論文以 Landsat 8 把這個轉移具體化：它每 22 秒看到一幀新影像，而一個實際的雲層過濾器每幀需要 98 秒，所以直接部署這個過濾器，下傳的高價值資料只增加 9%，理想過濾器卻能帶來三倍[83, §2.1.3, Fig. 5]。 |
| 星座配置 | 在 15 W 的 Jetson AGX Orin、140 W 的 Core i7-7800X 與 180 W 的 GTX 1070 Ti 上執行七個雲層過濾分割應用，並以 cote 模擬 Landsat 8 的軌道、相機與地面段，Kodan 把飽和下行鏈路的資料價值密度相對於彎管提高 89% 到 97%，並且相對於先前的 OEC 管線，把完整地面軌跡覆蓋所需的衛星數最多減少 12 倍[83, §4, §5.1–5.2, Figs. 8, 11]。 |
1. Where this paper sits1. 這篇論文的位置
Chapter 8 teaches computing in orbit as three walls that appear in sequence, and Kodan is the second wall. Denby and Lucia [60] identify the first bottleneck: a 200 Mbit/s station receives less data than the camera produces. On-board inference reduces that downlink demand. Kodan observes that the fix creates its own shortage.
The processor must fit satellite volume, mass, energy and cost limits while processing every frame. The study represents it with a 15 W Jetson AGX Orin. The camera keeps its original capture rate. Xing et al. [85] identify heat as the third bottleneck.
A few watts of sustained computation saturate small-satellite thermal paths before processor or radio limits. The three studies show the pattern repeated at larger scale in Part III. Relieving one resource limit exposes the next.
第 8 章把在軌運算教成三道依序出現的牆，Kodan 是第二道。第一道由同一位第一作者在 Denby 與 Lucia[60]記錄，是下行鏈路：200 Mbit/s 的地面站接收量低於奈米衛星相機產生的資料量，所以推論必須在機上進行。Kodan 觀察到這個解法製造了自己的短缺，因為現在必須對每一幀的每個圖塊跑分割網路的處理器，只能是塞得進衛星體積、質量、能源與成本預算的元件，論文以 15 W 模式的 Jetson AGX Orin 為代表，而相機仍依原速率產生影像。第三道牆是熱，隨 Xing 等人[85]而來，他們飛了商用板子，發現持續幾瓦的運算在觸及任何處理器或無線電限制之前，就已經讓小衛星的散熱路徑飽和。三篇合在一起讀，呈現出本教材第三部分在更大尺度上反覆出現的模式：每解除一種資源的限制，就暴露出下一種。
Kodan connects CMU orbital-edge research with Microsoft Research space systems. The latter later produced FedSpace, reviewed after Chapter 9. The shared model treats satellites as resource-limited nodes with intermittent connectivity. An onboard runtime allocates scarce resources between ground contacts.
Kodan 也是 CMU 軌道邊緣路線通往 Microsoft Research 太空系統興趣的橋樑，後者之後產出了第 9 章後導讀的 FedSpace 聯邦學習框架。共同前提是：衛星是一個弱小、間歇連線的節點，其稀缺資源必須由執行期自動來分配。
2. The problem, measured2. 問題，用量測說話
Kodan measures a mismatch between frame-arrival and compute clocks in three steps. The cote simulator models orbit, sensing, communication and ground infrastructure. Published Landsat performance validates the model. [83, §2.1]Kodan 量的是兩個時鐘之間的差距，論文分三步用 cote 模擬器把它建立起來。cote 模擬軌道、感測、通訊與地面段，作者以公開的 Landsat 效能數字驗證過它[83, §2.1]。第一步是下行鏈路：一圈軌道內，Landsat 地面段只能接收 2% 可得的萬像素高光譜觀測。而要每天觀測到每一個 Landsat 幀，星座至少需要 40 顆衛星[83, §2.1.1, Figs. 2–3]。成本的錨點則是 Landsat 任務的 8.55 億美元，作為單體衛星動輒數億美元的例子[83, §2]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Constellation configuration | The Landsat ground segment receives only 2% of available hyperspectral observations over one orbit. Frames have 10,000 pixels. Adding satellites raises downlinked frames from 5 with one satellite to 60 with sixteen, until stations saturate. A constellation needs at least 40 satellites to observe every Landsat frame each day [83, §2.1.1, Figs. 2–3] . |
| Accuracy | Cloud obscures about 67% of images. Of nearly 3,600 daily observable Landsat 8 frames, only a third are high-value. A bent pipe delivers less than 21% of observable high-value data. An ideal on-board filter with perfect accuracy and zero execution time would deliver 63%, three times as much [83, §2.1.2, Fig. 4] . |
| Evaluation conditions | The third step is the processor. A cloud filter takes 98 s against a 22 s frame deadline. It processes only a fraction of frames, raising high-value downlink by 9%. [83, §2.1.3, Fig. 5] |
| Onboard processors | Related work lists the 32-bit SPARC V8 LEON at 250 MHz, EO-1 Mongoose-V at 12 MHz, and RAD5500 at 466 MHz. After two or more decades, these processors "could be nearly half a century behind the state-of-the-art". The $855 million Landsat mission illustrates monolithic satellite costs of hundreds of millions of dollars. [83, §6] [83, §2] |
| 軌道面配置 | 在同一軌道面增加衛星，可以把每個週期下傳的幀數從一顆衛星的 5 幀提高到十六顆的 60 幀，但只到地面站飽和為止。 |
| 影像篩選收益 | 第二步是過濾的價值：約 67% 的衛星影像被雲遮蔽，所以一顆 Landsat 8 一天可觀測的近 3,600 幀中只有三分之一是高價值的，彎管只送下這些可觀測高價值資料的低於 21%，而準確率完美、執行時間為零的理想機上過濾器可送下 63%，是三倍[83, §2.1.2, Fig. 4]。 |
| 評估條件 | 第三步是處理器：一個實際的雲層過濾器每幀需要 98 秒，而每幀期限是 22 秒，所以直接部署只能過濾部分幀，下傳的高價值資料只增加 9%[83, §2.1.3, Fig. 5]。 |
| 評估條件 | 構成問題背景的飛行履歷元件出現在相關研究一節，作者列出 LEON（達到 250 MHz 的 32 位元 SPARC V8 核心）、EO-1 上 12 MHz 的 Mongoose-V 與 466 MHz 的 RAD5500，並指出運作二十年以上之後，這類處理器「可能落後最先進技術近半個世紀」[83, §6]。 |
Data-value density measures the high-value share of bytes crossing a saturated downlink. For example, land pixels with clear skies provide value while ocean cloud tops may add little. A bent pipe downlinks whatever the camera saw and has the lowest density. An ideal on-board filter would downlink only valuable pixels. And a compute-bottlenecked satellite sits between them because it can filter only the frames it has time to process.
論文自己衡量利害的指標是資料價值密度，也就是通過飽和下行鏈路的位元組中，攜帶任務想要資訊的比例，例如陸地上晴空陸地像素，海洋雲頂則屬於低價值內容。彎管把相機看到的全部下傳，密度最低。理想的機上過濾器只下傳有價值的像素。運算受限的衛星介於兩者之間，因為它只能過濾來得及處理的那些幀。
3. The design3. 設計
Kodan's design is a one-time transformation before launch followed by a runtime that applies it [83, §3.1, Fig. 7] . Before launch, Kodan partitions representative data into contexts. Experts can define contexts such as ocean, desert and city. Position and orientation then locate the context on a map.
Alternatively, k-means groups label vectors containing geographic and weather features. A small classification network, the context engine, labels orbital tiles. The model treats its output as ground truth. [83, §3.2] It trains models for individual contexts or context groups. Each has measured target-hardware execution time and precision.
A specialized model can be smaller than the datacenter reference application while keeping or improving its accuracy [83, §3.3] . A cote simulation sweeps tiles per frame and processing elision for each context. It models target orbit, sensor, radio and ground infrastructure. The combination maximizing data-value density becomes the selection logic. [83, §3.4]
Kodan 的設計是發射前的一次性轉換，加上在軌套用它的執行期[83, §3.1, Fig. 7]。發射前，Kodan 把代表性資料集分成情境，可以由領域專家劃分，此時海洋、沙漠或城市之類的情境可由衛星位置與姿態投影到地圖上得知。也可以自動劃分，以 k-means 聚類記錄各樣本地理與天氣特徵的標籤向量，此時由一個小型分類網路，也就是情境引擎，在軌為每個圖塊標上情境，並把它的輸出當作真值[83, §3.2]。接著它訓練特化於單一情境或一組情境的模型，每個模型在目標硬體上的執行時間與精確度都已知，而特化模型可以比資料中心的參考應用更小，同時維持或提高準確率[83, §3.3]。最後，它對目標衛星的軌道、感測器、無線電與地面段做 cote 模擬，掃描每幀圖塊數與各情境的處理或略過選項，把使資料價值密度最大的組合保留為選擇邏輯[83, §3.4]。
At runtime, frames are divided into the selected tile count. The context engine labels each tile. Selection logic chooses raw downlink, discard or a context-specialized model. [83, §3.1] Tiling is the time knob. A 10,000 × 10,000-pixel frame becomes four 5,000-pixel tiles or sixteen 2,500-pixel tiles.
Each tile is resized to network input dimensions. Thus, the tile count sets the frame's processing time while fewer, larger tiles lose detail to decimation [83, §3, Fig. 6] . When mean runtime exceeds the frame deadline, Kodan trades precision for speed through coarser tiling and high-value-context elision. When runtime fits, idle time supports more precise specialized models.
Elision then targets contexts containing almost entirely high-value data. [83, §3.4]
在軌時，每一幀被切成選定數目的圖塊，情境引擎為每個圖塊標上情境，選擇邏輯再決定這個圖塊是直接下傳原始內容、丟棄，還是用其情境的特化模型處理[83, §3.1]。切塊是時間的旋鈕：一幀 10,000 × 10,000 像素的 Landsat 影像切成四塊 5,000 像素或十六塊 2,500 像素的圖塊，再逐塊縮放到網路的輸入尺寸，所以圖塊數決定了每幀的處理時間，而較少、較大的圖塊會因降取樣而失去細節[83, §3, Fig. 6]。策略跟著瓶頸走：應用平均超出每幀期限時，Kodan 以較粗的切塊與略過高價值情境，用較低精確度換取較短執行時間。趕得上期限時，它把閒置時間花在最精確的特化模型上，只略過幾乎全是高價值資料的情境[83, §3.4]。
Original figure note原始圖說出處
Drawn for this primer from Denby et al., ASPLOS 2023, §3, Figs. 6–7.4. Evaluation setup4. 實驗設定
The evaluation combines execution times measured on ground hardware with a simulated orbit, with flight validation outside its scope [83, §4] . Landsat 8 supplies orbit, camera and World Reference System ground-track captures. Ground stations retain Landsat positions and link characteristics. cote derives each deployment's frame deadline from orbit. Thus, the downlink is a simulated, saturated Landsat ground segment rather than a rate the paper states.
The design uses local computation and ground links only. The two baselines are the bent pipe and the direct deployment of each unmodified application [83, §4, §5.1] .
評估結合了在地面硬體上量測的執行時間與模擬的軌道，飛行驗證則位於範圍之外[83, §4]。軌道、相機與影像擷取都是 Landsat 8 沿其世界參考系統（WRS）地面軌跡的情形，地面段是 Landsat 的地面站及其位置與鏈路特性，cote 由軌道算出每個部署的每幀期限，所以下行鏈路是模擬的、飽和的 Landsat 地面段，其設定依據為模擬的地面段。星間鏈路位於設計範圍之外，兩個基準是彎管與直接部署原版應用[83, §4, §5.1]。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Training comparison | The input is the Sentinel-2 Cloud Mask Catalogue. Its samples are 48% high-value (cloud-free) and 52% low-value, with a subset held out for validation and data augmentation during training. |
| Evaluation platforms | The seven applications are publicly available semantic-segmentation networks customised to output a per-pixel cloud mask, from mobilenetv2dilated-c1-deepsup as Application 1 to resnet101dilated-ppm-deepsup as Application 7. The Jetson AGX Orin runs at 15 W, near the authors' reasonable upper limit for a 3U CubeSat subsystem. The 12-core i7-7800X runs at 3.5 GHz and about 140 W. The GeForce GTX 1070 Ti draws about 180 W. Application 1 takes 178 ms per tile on the 1070 Ti. Application 7 takes 2,545 ms on the i7. [83, §4, Tab. 1] |
| 訓練比較 | 輸入是 Sentinel-2 Cloud Mask Catalogue，樣本中 48% 為高價值（晴空）、52% 為低價值，保留一部分作驗證，訓練時並做資料擴增。 |
| 評估運算平台 | 七個應用都是公開的語意分割網路，改成輸出逐像素雲遮罩，從應用 1 的 mobilenetv2dilated-c1-deepsup 到應用 7 的 resnet101dilated-ppm-deepsup，每個都在三個平台上執行：15 W 模式的 NVIDIA Jetson AGX Orin（作者認為接近 3U 立方衛星子系統合理功耗的上限）、12 核 3.5 GHz、約 140 W 的 Core i7-7800X，以及約 180 W 的 GeForce GTX 1070 Ti。 |
| 評估條件 | 每個圖塊的處理時間從應用 1 在 1070 Ti 上的 178 ms 到應用 7 在 i7 上的 2,545 ms[83, §4, Tab. 1]。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| Data-value density of a saturated downlink improves by 89–97% over the bent pipe across all seven applications and all three platforms [83, §5.1, Fig. 8].資料價值密度相對於彎管提高 89–97%，涵蓋全部七個應用與三個平台[83, §5.1, Fig. 8]。 | The bent pipe's density equals the dataset's 48% high-value share, so the gain corresponds to Kodan densities of roughly 0.91–0.95 (calc table). The constrained 15 W Orin gains most. Direct Application 1 deployment on the 1070 Ti already meets the deadline.彎管的密度等於資料集 48% 的高價值比例，所以這個增益對應 Kodan 約 0.91–0.95 的密度（見計算表）。增益在最受限的 15 W Orin 上最大，而在 1070 Ti 上，直接部署應用 1 已能趕上期限。 |
| Up to 12× fewer satellites are needed for full ground-track coverage than with direct deployment and prior OEC pipelining [83, §5, Fig. 11].達到完整地面軌跡覆蓋所需的衛星數，比直接部署加上先前的 OEC 管線最多少 12 倍[83, §5, Fig. 11]。 | A consequence of shorter per-frame processing time, computed in simulation against the OEC baseline that spreads frames across a pipeline of satellites. Constellation flight validation remains pending.這是每幀處理時間縮短的結果，在模擬中相對於把影像分散到一串衛星管線的 OEC 基準計算而得。星座飛行驗證仍待完成。 |
| Context specialization raises precision by 33% for Application 2 and accuracy by up to 7.5% [83, §5.3, Fig. 12].情境特化讓應用 2 的精確度提高 33%，準確率最多提高 7.5%[83, §5.3, Fig. 12]。 | Precision here is the fraction of correctly labelled pixels, which sets density on a saturated link. The applications with the lowest baseline accuracy gain most.這裡的精確度是正確標記像素的比例，在飽和鏈路上決定密度。基準準確率最低的應用增益最大。 |
| Tiling gives up to 50% more density for Application 7 on the Orin, where nine tiles per frame beat 16, 36 and 121 [83, §5.4, Figs. 13–14].切塊讓應用 7 在 Orin 上的密度最多提高 50%，每幀九塊勝過 16、36 與 121 塊[83, §5.4, Figs. 13–14]。 | On the 1070 Ti, where the bottleneck eases, the tiling that maximizes precision also maximizes density, and the precision-optimal tile count differs by architecture.在瓶頸較鬆的 1070 Ti 上，使精確度最大的切塊也使密度最大，而精確度最佳的圖塊數因架構而異。 |
| Context-based elision raises density for Application 1 by 39% on the Orin and 34% on the i7 [83, §5.5, Fig. 15].依情境略過處理讓應用 1 的密度在 Orin 上提高 39%、在 i7 上提高 34%[83, §5.5, Fig. 15]。 | The 34% is an elision result, a distinct quantity from Kodan's overall i7 gain. On the 1070 Ti Application 1 meets its frame deadline, and elision helps only by sending likely high-value pixels unprocessed rather than imprecisely filtered.34% 是略過處理的結果，與 Kodan 在 i7 上的整體增益各有定義。在 1070 Ti 上應用 1 能趕上每幀期限，略過處理的幫助只在於把可能是高價值的像素直接送下原始內容，以減少低精確度過濾造成的損失。 |
| Flight-heritage processors run at 12 MHz (Mongoose-V), 250 MHz (LEON) and 466 MHz (RAD5500) [83, §6].飛行履歷處理器的時脈為 12 MHz（Mongoose-V）、250 MHz（LEON）與 466 MHz（RAD5500）[83, §6]。 | Motivating numbers from the related-work section. Kodan was evaluated on an Orin, an i7 and a 1070 Ti, none of them flight hardware.來自相關研究一節的動機數字。Kodan 是在 Orin、i7 與 1070 Ti 上評估的，三者皆屬於地面評估硬體。 |
| Question問題 | Formula公式 | Result結果 |
|---|---|---|
| What multiplier does a 89–97% improvement in density correspond to?密度提高 89–97% 相當於多少倍？ | \( 1 + 0.89 \le \dfrac{\rho_{\text{Kodan}}}{\rho_{\text{bent}}} \le 1 + 0.97 \) | 1.89–1.97×computed from the percentage range of §5.1由 §5.1 的百分比範圍計算 |
| What densities does that range imply on the paper's dataset?在論文的資料集上，這個範圍隱含多少密度？ | \( \rho_{\text{bent}} = 0.48,\quad \rho_{\text{Kodan}} = (1.89\text{–}1.97) \times 0.48 \) | ≈ 0.91–0.95computed. 48% high-value share from §4, bent-pipe density equal to that share per §5.1. Consistent with the Kodan bars of Fig. 8計算值。48% 高價值比例取自 §4，依 §5.1 彎管密度等於該比例。與 Fig. 8 中 Kodan 的長條一致 |
| What share of frames can a 98 s filter finish within a 22 s deadline?98 秒的過濾器能在 22 秒期限內處理完多少比例的幀？ | \( 22 / 98 \) | ≈ 22%computed from §2.1.3. The paper reports the resulting gain in high-value data downlinked as 9%由 §2.1.3 計算。論文報告由此得到的高價值資料下傳增益為 9% |
| Per-frame time of Application 7 on the i7 at nine tiles, with elision disabled應用 7 在 i7 上每幀切九塊、完整執行處理時的每幀時間 | \( 9 \times 2.545\ \text{s} \) | ≈ 22.9 scomputed from Table 1. Just above a 22 s deadline, which is why elision still matters on the i7由 Table 1 計算。略高於 22 秒期限，所以在 i7 上略過處理仍然重要 |
| How far behind is a 466 MHz RAD5500 against a 250 MHz LEON?466 MHz 的 RAD5500 與 250 MHz 的 LEON 差多少？ | \( \dfrac{466}{250} \approx 1.9 \) | ≈ 1.9× clockcomputed from §6. Clock alone, the paper leaves throughput figures unspecified for either part由 §6 計算。僅時脈，論文各元件吞吐量皆仍待補充 |
6. Limits and open questions6. 限制與待解的問題
- Ground hardware and a simulated orbit. The motivation cites 12–466 MHz flight-heritage processors. Evaluation instead combines Orin, i7 and 1070 Ti execution times with a Landsat 8 cote simulation. The 15 W Orin is closest to CubeSat conditions. Radiation tests and flight validation remain pending for all three evaluation platforms. Thus, the single-event upsets and flight-processor behavior require separate evidence.地面硬體加上模擬軌道。瓶頸以 12–466 MHz 的飛行履歷元件來鋪陳，示範卻是把在 Orin、i7 與 1070 Ti 上量到的執行時間，結合 Landsat 8 的 cote 模擬。三者中 15 W 的 Orin 最接近立方衛星，但三者的輻射測試與飛行驗證皆仍待完成，所以單事件翻轉與飛行處理器表現皆須另行取得證據。
- Dataset and orbit come from separate missions. The samples are Sentinel-2 images while the frames, deadlines and ground segment are Landsat 8's. The dataset's 48% high-value share, with the cited 67% global cloud fraction describing a separate quantity the paper cites in §2, sets the bent-pipe baseline. A cloudier distribution would lower the bent-pipe density and change how much elision gains.資料集與軌道差異。樣本是 Sentinel-2 影像，而影像幀、期限與地面段卻是 Landsat 8 的。決定彎管基準的是資料集 48% 的高價值比例，論文 §2 引用的全球 67% 雲量則屬於另一項統計。雲更多的分布會降低彎管密度，也會改變略過處理的增益。
- Thermal budget absent. Kodan assumes a fixed compute capacity per frame. Xing et al. find a Raspberry Pi 4B above 75% load reaches throttling within about 50 minutes in orbit. Warming therefore reduces achievable runtime performance. Kodan leaves this effect outside its scheduler model. [85]熱預算位於模型範圍之外。Kodan 假設每幀的運算容量固定。Xing 等人[85]發現 Raspberry Pi 4B 在負載超過 75% 時約 50 分鐘內就在軌進入降頻區，所以執行期該據以規劃的期限會隨衛星升溫而縮短，Kodan 的排程器將此效應排除於模型之外。
- Saturated downlink by assumption. The data-value-density metric only makes sense when the downlink is the binding constraint. Chapter 3 records 200 Gbps-class TBIRD downlinks. At that rate, raw-data bent pipes may again suit some missions. This would change the comparison baseline.下行飽和是假設。資料價值密度這個指標只在下行鏈路是束縛限制時才有意義。第 3 章顯示 TBIRD 已飛過 200 Gbps 等級的光學下行，在那種速率下某些任務的彎管可能重新變得有吸引力，比較基準也會隨之改變。
- Local processing and ground links only. The decision is per satellite and per tile. Neighbor-assisted processing remains open. OEC coordinates work through formation, while Kodan leaves ISL rate, latency and topology outside its scope.星間鏈路位於設計範圍之外。決定是逐衛星、逐圖塊。過載衛星依 OEC 編隊排程把區塊交給鄰近衛星的可行性，仍待探討，所以 Kodan 將 ISL 速率、延遲及拓樸列於範圍之外。
- The context engine is taken as ground truth. For automatically generated contexts the engine's output is treated as ground truth when the dataset is partitioned [83, §3.2]. The evaluation leaves unreported how often a tile's context is misclassified or what that costs in valuable tiles discarded. Cloud changes hourly, so map-based expert contexts would face the same question.情境引擎被當作真值。自動產生情境時，劃分資料集就把引擎的輸出當作真值[83, §3.2]，評估圖塊情境誤分類的頻率及因此丟棄的高價值圖塊量皆仍待報告。雲逐小時變化，所以依地圖劃分的專家情境也會遇到同樣的問題。
- Data-value density within the transmitted stream. Data-value density divides high-value downlinked bytes by total downlinked bytes. Discarded or misclassified high-value tiles lie outside that denominator. The evaluation in §5 reports how clean the downlinked stream is rather than how much of the observable high-value data reaches the ground. This is the quantity §2 used to motivate the work.指標衡量傳輸資料流內的價值密度。資料價值密度是通過飽和鏈路的資料中高價值的比例，所以被略過處理丟掉或被過濾器標錯的高價值圖塊位於其分母範圍之外。§5 的評估報告的是下傳資料流有多乾淨，可觀測高價值資料的總送達量則屬於另一項指標，後者正是 §2 用來說明研究動機的量。
- Seven applications of one kind. All seven workloads are cloud-filtering segmentation, so the 89–97% range speaks to one task family. Detection or change-tracking workloads with different value distributions could behave differently.七個同類應用。七個工作負載全是雲過濾分割，所以 89–97% 的範圍只對一類任務成立。價值分布相異的偵測或變化追蹤工作負載可能表現各異。
7. How it connects7. 與其他內容的關聯
- Section 8.1: the 89–97% figure is the second step of the downlink-wall argument, after OEC's 112→6 stations.第 8.1 節：89–97% 是下行瓶頸論證的第二步，接在 OEC 的 112→6 座地面站之後。
- Section 8.3: heat is the constraint Kodan's deadline model leaves out, and the chapter's Pi and Atlas curves show how fast it bites.第 8.3 節：熱是 Kodan 期限模型漏掉的限制，該章的 Pi 與 Atlas 曲線顯示它多快發作。
- Section 8.4: why flight processors lag, with the radiation table that explains the LEON and RAD5500 choice.第 8.4 節：飛行處理器為何落後，附有解釋 LEON 與 RAD5500 選擇的輻射表。
- Section 3.6: the Earth-observation downlink numbers that make the downlink saturated in the first place.第 3.6 節：讓下行鏈路一開始就飽和的地球觀測下行數字。
- Section 11.4: the same "specified versus open" reading applied to a proposal whose compute budget is three orders of magnitude larger.第 11.4 節：把同樣「已指明與待指明」的讀法用在一個運算預算大三個數量級的提案上。
- OEC [60] is the first wall Kodan builds on. Xing et al. [85] supply the thermal budget Kodan lacks. Bhattacherjee et al. [2] imagine a 64-core server where Kodan evaluates a 15 W Orin. FedSpace [91] continues the Microsoft Research line into training.OEC[60]是 Kodan 建立於其上的第一道牆。Xing 等人[85]提供 Kodan 缺少的熱預算。Bhattacherjee 等人[2]在 Kodan 評估 15 W Orin 的地方想像一台 64 核伺服器。FedSpace[91]把 Microsoft Research 這條路線延伸到訓練。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| +89–97% | Data-value density of a saturated downlink vs the bent pipe, seven applications on three platforms [83, §5.1, Fig. 8] | 飽和下行鏈路的資料價值密度對彎管，七個應用、三個平台[83, §5.1, Fig. 8] | |
| ≈ 0.48 → 0.91–0.95 | Bent-pipe density (the dataset's high-value share) and the Kodan density the gain implies (computed) [83, §4, §5.1] | 彎管密度（資料集的高價值比例）與增益隱含的 Kodan 密度（計算值）[83, §4, §5.1] | |
| 22 s · 98 s | Landsat 8 frame deadline and a real cloud filter's time per frame [83, §2.1.3] | Landsat 8 每幀期限與實際雲層過濾器每幀所需時間[83, §2.1.3] | |
| +9% vs 3× | High-value data downlinked: direct deployment of the filter vs an ideal filter [83, §2.1.2–2.1.3] | 下傳的高價值資料：直接部署過濾器對理想過濾器[83, §2.1.2–2.1.3] | |
| up to 12× | Fewer satellites for full ground-track coverage than direct deployment with OEC pipelining [83, §5, Fig. 11] | 完整地面軌跡覆蓋所需衛星數比直接部署加 OEC 管線少的倍數[83, §5, Fig. 11] | |
| +39% · +34% | Gain from context-based elision for Application 1 on the Orin and on the i7 [83, §5.5] | 依情境略過處理讓應用 1 在 Orin 與 i7 上的增益[83, §5.5] | |
| up to +50% | Gain from tiling for Application 7 on the Orin (nine tiles per frame) [83, §5.4] | 切塊讓應用 7 在 Orin 上的增益（每幀九塊）[83, §5.4] | |
| +33% · ≤ 7.5% | Precision gain of Application 2 and accuracy gain from geospatial contexts [83, §5.3] | 應用 2 的精確度增益與地理情境帶來的準確率增益[83, §5.3] | |
| 15 W · 140 W · 180 W | Jetson AGX Orin, Core i7-7800X and GTX 1070 Ti evaluation platforms [83, §4] | 評估平台 Jetson AGX Orin、Core i7-7800X 與 GTX 1070 Ti[83, §4] | |
| 250 · 12 · 466 MHz | LEON, Mongoose-V (EO-1) and RAD5500 flight-heritage processors [83, §6] | LEON、Mongoose-V（EO-1）與 RAD5500 飛行履歷處理器[83, §6] | |
| ≈ 50 years | How far behind the state of the art an on-board processor can be after two or more decades in service, the authors' phrase [83, §6] | 機上處理器服役二十年以上後可能落後最先進技術的程度，作者的說法[83, §6] | |
| $855M | Landsat mission cost, quoted as an example of a monolithic satellite [83, §2] | Landsat 任務成本，作為單體衛星的例子[83, §2] |