Electronic Theses and Dissertation

Infrastruktur Software Defined Network (SDN) memisahkan lapisan kendali dari lapisan penerusan data sehingga lalu lintas jaringan dapat dikelola secara terpusat dan terprogram. Pada lingkungan SDN yang dinamis, load balancing berperan kritis dalam menjaga distribusi beban yang proporsional terhadap kapasitas server, terutama ketika terjadi perubahan bobot alokasi. Penelitian ini mengusulkan strategi distribusi proporsional tiga fase yang membangun justifikasi matematis, validasi empiris, dan kerangka adaptif berbasis pembelajaran penguatan mendalam, dengan metode Sainte-Laguë (SL) sebagai komponen inti yang diadaptasi dari alokasi kursi parlemen ke distribusi beban jaringan. Penelitian ini dilakukan pada 3 fase, fase pertama memvalidasi keunggulan proporsionalitas SL secara matematis. Diukur dengan Mean Absolute Percentage Error (MAPE), SL mencapai rata-rata 1,8%, jauh di bawah Weighted Round Robin (WRR) sebesar 8,3% dan Interleaved Weighted Round Robin (IWRR) sebesar 6,1%. Fase kedua mengevaluasi empat algoritma (SL, WRR, IWRR, dan Weighted Least Connection (WLC)) pada emulasi Mininet melalui tiga skenario. Pada kondisi stasioner keempatnya setara, tetapi setelah perubahan bobot hanya SL yang konvergen berkelanjutan dengan MAPE akhir 4,8%, sedangkan WRR, IWRR, dan WLC gagal kembali ke distribusi ideal. Fase ketiga membangun PPO-SL-QoS, framework two-timescale yang mengintegrasikan agen PPO dengan SL pada topologi Fat-Tree k=4. Kerangka ini menjamin proteksi kelas Gold di atas target pada semua kondisi perubahan bobot (bw_ratio_gold 1,20–1,40) dan unggul penuh atas lima baseline dengan Cliff's delta 1,000. Hasil ini mengonfirmasi bahwa jaminan proteksi berbasis arsitektur lebih unggul dibanding pendekatan berbasis pelatihan semata.

Software-Defined Networking (SDN) infrastructure decouples the control plane from the data forwarding plane, allowing network traffic to be managed cen-trally and programmatically. In dynamic SDN environments, load balancing plays a critical role in maintaining a proportional load distribution relative to server capaci-ty, particularly when allocation weights change. This study proposes a three-phase proportional distribution strategy that establishes mathematical justification, empiri-cal validation, and an adaptive framework based on deep reinforcement learning, with the Sainte-Laguë (SL) method as its core component, adapted from parliamen-tary seat apportionment to network load distribution. The research is conducted in three phases. The first phase mathematically validates the proportionality advantage of SL. Measured by the Mean Absolute Percentage Error (MAPE), SL achieves an average of 1.8%, far below Weighted Round Robin (WRR) at 8.3% and Interleav-ed Weighted Round Robin (IWRR) at 6.1%. The second phase evaluates four algo-rithms (SL, WRR, IWRR, and Weighted Least Connection (WLC)) in a Mininet emulation across three scenarios. Under stationary conditions the four perform comparably, but after weight changes only SL achieves sustained convergence with a final MAPE of 4.8%, while WRR, IWRR, and WLC fail to return to the ideal dis-tribution. The third phase develops PPO-SL-QoS, a two-timescale framework that integrates a PPO agent with SL on a Fat-Tree k=4 topology. The framework guar-antees Gold-class protection above target under all weight-change conditions (bw_ratio_gold 1.20–1.40) and fully outperforms five baselines with a Cliff's delta of 1.000. These results confirm that architecture-guaranteed protection is superior to approaches based on training alone.