Context Aware Provisioning Dan Incremental Snapshotting Pada Workspace Stateful LLM Agent Menggunakan Firecracker MicroVM

Fraditya, Awang (2026) Context Aware Provisioning Dan Incremental Snapshotting Pada Workspace Stateful LLM Agent Menggunakan Firecracker MicroVM. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5027221055-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Perkembangan Large Language Models (LLM) menjadi agen otonom menuntut lingkungan eksekusi yang mampu mempertahankan status (stateful) guna menyimpan konteks percakapan dan dependensi alat. Namun, infrastruktur serverless saat ini memiliki keterbatasan berupa latensi cold-start yang tinggi dan sifat stateless, sedangkan solusi berbasis Virtual Machine persisten tidak efisien dalam penggunaan sumber daya. Penelitian ini mengusulkan arsitektur pengelolaan workspace Agen LLM berbasis Firecracker MicroVM yang menerapkan pendekatan Context-Aware Provisioning dan Incremental Snapshotting dengan mekanisme Copy-on-Write (CoW) berlapis serta Garbage Collection (GC) berbasis compaction pada batas max_chain_length=5. Hasil pengujian menunjukkan bahwa Incremental Snapshotting berhasil mengurangi ukuran unggahan snapshot secara substansial sebesar 97-99% pada beban kerja idle dan light (dari 92-95 MB menjadi 0-3 MB) serta mempercepat proses pembuatan hingga 20,1×. Pada beban kerja heavy dan extreme, reduksi ukuran mencapai 33-45%. Meskipun demikian, terdapat trade-off berupa latensi resume yang 1,26-1,83× lebih lambat akibat proses dekompresi yang menyumbang 56-80% bottleneck dan penambahan level rantai yang linier sebesar 2,5 detik per tingkat. Namun, akumulasi penyimpanan selama 7 snapshot dengan GC terbukti jauh lebih efisien (417 MB) dibandingkan full snapshot (822 MB). Sementara itu, Context-Aware Provisioning memberikan peningkatan efisiensi yang adaptif, mempercepat ready time sebesar 21,8% pada lingkungan Python (uv), namun pada lingkungan Frontend (bun), fresh boot justru lebih cepat 614 ms. Pemulihan dari snapshot terbukti hanya menguntungkan jika durasi instalasi dependensi melebihi 8 detik. Pada pengujian end-to-end (E2E) non-deterministik, eliminasi satu langkah loop ReAct melalui penyediaan lingkungan yang tepat mampu menghemat satu siklus inferensi penuh (~2–5 detik) dan memangkas konsumsi token sebesar 12,4-19,4%, memberikan dampak multiplikatif yang signifikan terhadap responsivitas sistem dari perspektif pengguna akhir.
======================================================================================================================================
The evolution of Large Language Models (LLMs) into autonomous agents requires stateful execution environments to maintain conversation context and tool dependencies. However, current serverless infrastructures suffer from high cold-start latency and are inherently stateless, while persistent Virtual Machine solutions are resource-intensive and cost-inefficient. This study proposes a workspace management architecture for LLM agents using Firecracker MicroVMs, incorporating Context-Aware Provisioning and Incremental Snapshotting via multi-layered Copy-on-Write (CoW) mechanisms and compaction-based Garbage Collection (GC) restricted to max_chain_length=5. The experimental results demonstrate that Incremental Snapshotting substantially reduces snapshot upload sizes by 97-99% under idle and light workloads (from 92-95 MB to 0.3-3 MB) while accelerating creation speeds by up to 20.1×. Under heavy and extreme workloads, size reductions range between 33-45%. However, a trade-off is observed as resume latency is 1.26-1.83× slower than full snapshots, primarily due to decompression accounting for 56-80% of the bottleneck and an additional 2.5 seconds added linearly per chain level. Nevertheless, cumulative storage usage over 7 snapshots with GC is significantly more efficient (417 MB) compared to full snapshots (822 MB). Furthermore, Context-Aware Provisioning delivers adaptive efficiency, improving ready times by 21.8% in Python (uv) environments, whereas a fresh boot is 614 ms faster in Frontend (bun) environments. Restoring from a snapshot proves advantageous only when dependency installation exceeds 8 seconds. In non-deterministic end-to-end (E2E) testing, eliminating a single ReAct loop step by providing the appropriate pre-configured environment circumvents an entire LLM inference cycle (~2-5 seconds) and decreases token consumption by 12.4-19.4%, yielding a substantial multiplicative effect on overall system responsiveness for the end-user.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	MicroVM, Firecracker, Incremental Snapshotting, Context-Aware Provisioning, AI Agent
Subjects:	T Technology > T Technology (General) > T57.62 Simulation T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information Technology > 59201-(S1) Undergraduate Thesis
Depositing User:	Awang Fraditya
Date Deposited:	18 Jul 2026 05:06
Last Modified:	18 Jul 2026 05:06
URI:	http://repository.its.ac.id/id/eprint/135225

Actions (login required)

View Item