邊緣運算環境中，串流型或時間敏感型服務須在使用期間持續接收請求並於指定時間內回傳結果，僅在回應延遲與更新頻率皆符合各自之服務品質（Quality of Service， QoS）門檻時，該服務方視為對使用者有效；當使用者同時依賴多項服務時，任一服務未達門檻即無法取得完整應用功能。受限於邊緣節點算力有限，當使用者數量動態變化、甚至節點故障時，系統須在資源受限下動態調整服務部署與資源配置，以維持更多使用者之完整服務。然而，現有研究的優化目標多以延遲與成本加權最小化（如EMS）、延遲硬約束下成本最小化（如SPARQ）、或覆蓋率與負載均衡最大化（如LSR）為設計邏輯，均未以「QoS 門檻達標人數」為優化對象，因此在資源不足時無法對「該優先保障誰」做出取捨。加權最小化目標傾向將資源平均分配給所有使用者，延遲硬約束目標在超載時僅能宣告問題不可行，覆蓋導向目標則以全有或全無之指派使部分訂閱服務遭結構性排除，三者最終皆可能使大量使用者落於可用門檻之下，形成無效服務。針對上述缺口，本研究提出服務部署與資源配置聯合優化的數學模型：以離線量測所得之服務容量建立容量約束，並以二元指示變數之連乘積將多項服務須同時達標之要求代數化為多服務共同達標條件，以最大化QoS 門檻達標之使用者人數為優化目標。本研究設計兩層式求解架構：上層以遺傳演算法搜尋部署配置，下層以難易序貪婪前瞻分派演算法決定各使用者請求速率；並以此數學模型為決策依據，建立快照比較後之准入機制，以及銜接靜態最佳解與動態冷啟動、分散式非同步執行之配置執行與狀態轉換機制，使新使用者可於冷啟動期間以降級但達標之速率提早加入，降低冷啟動期間服務中斷之風險。本研究於9台Amazon Web Services（AWS）Elastic Compute Cloud（EC2）構成之Kubernetes叢集上，以多使用者同時使用軟即時邊緣AI推論服務之情境進行實驗，與EMS、LSR兩個文獻移植基準比較，於動態日常、高負載、節點故障三種情境下驗證。結果顯示，本研究方法之主要指標（QoS達標人數×服務時間）於三種情境下均優於兩基準，超載情境下領先EMS、LSR分別達60%、66%，次要指標（服務品質×服務時間）亦分別領先43%、64%，顯示以QoS門檻達標為核心目標、並以實測容量建模之調度策略，能在資源受限下更有效維持邊緣AI推論服務之服務品質。;In edge computing environments， streaming or soft real-time services must continuously receive requests and return results within a specified time during their usage; a service is considered effective for a user only when both its response latency and update rate meet their respective Quality of Service (QoS) thresholds. When a user depends on multiple such services， failing to meet any one threshold prevents the user from obtaining the complete application functionality. Given the limited capacity of edge nodes， when the number of users changes dynamically or a node fails， the system must dynamically adjust service deployment and resource allocation under constrained resources to keep more users fully served. Existing optimization models (e.g.， EMS， SPARQ， LSR) do not target the number of users meeting a QoS threshold， and therefore cannot decide whom to prioritize under resource scarcity. Weighted-sum objectives tend to distribute resources evenly， hard-latency-constraint objectives can only declare infeasibility under overload， and coverage-oriented objectives structurally exclude partial subscriptions through all-or-nothing assignment; all three may ultimately leave many users below the usable threshold， resulting in ineffective service. To bridge this gap， this study proposes a joint optimization model for service deployment and resource allocation: offline-measured service capacity is used to construct a capacity constraint， while a product of binary indicator variables algebraically encodes the requirement that multiple services be satisfied together， jointly maximizing the number of users meeting the QoS threshold.
We design a two-tier framework: the upper tier employs a genetic algorithm for deployment， while the lower tier utilizes a difficulty-ordered greedy look-ahead allocation algorithm for request rates. Building on this model as a decision basis， we further design a snapshot-comparison-based admission mechanism and a configuration execution and state-transition mechanism that bridges the gap between the static optimum and dynamic cold-start and distributed asynchronous execution， reducing the risk of service disruption during cold start. Tested on a Kubernetes cluster with nine AWS EC2 instances under a multi-user soft real-time edge AI inference scenario， our approach was compared against EMS and LSR under normal， overloaded， and node-failure scenarios. Specifically， in the overloaded scenario， our method led EMS and LSR by 60% and 66% in the primary metric (QoS-compliant users x service time)， and by 43% and 64% in the secondary metric (service quality x service time). This shows that a strategy centered on QoS-threshold satisfaction， combined with empirical capacity modeling， more effectively maintains service quality for edge AI inference under tight resource constraints.
