Reproduksi Eksperimen dan Klaim dari ‘A Declarative System for Optimizing AI Workloads’

Suganda, Mathias Adya Diwangkara (2026) Reproduksi Eksperimen dan Klaim dari ‘A Declarative System for Optimizing AI Workloads’. Project Report. [s.n.], [s.l.]. (Unpublished)

Text 5025231015-Project_Report.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Peningkatan kapabilitas Large Language Models (LLM) telah mendorong munculnya kelas beban kerja baru yang dikenal sebagai Semantic Analytics Applications (SAPPs). Meskipun menawarkan kemampuan penalaran yang kuat pada data tidak terstruktur, implementasi SAPPs sering kali terkendala oleh biaya finansial yang tinggi dan latensi eksekusi yang lambat. Palimpzest hadir sebagai kerangka kerja deklaratif yang memisahkan spesifikasi logis program dari implementasi fisiknya, memungkinkan optimasi otomatis. Penelitian kerja praktik ini berfokus pada reproduksi eksperimen sistem Palimpzest dengan mengintegrasikan optimizer berbasis biaya terbaru, yaitu Abacus. Eksperimen dilakukan dengan menjalankan berbagai beban kerja standar, termasuk Legal Discovery dan Real Estate Search, untuk membandingkan rencana eksekusi yang dihasilkan oleh Abacus melawan implementasi baseline naif. Hasil pengujian menunjukkan bahwa integrasi Abacus dengan model-model baru mampu mengidentifikasi rencana eksekusi yang lebih efisien. Studi ini memvalidasi klaim pada paper.
=====================================================================================================================================
Advancements in the capabilities of Large Language Models (LLMs) have driven the emergence of a new class of workloads known as Semantic Analytics Applications (SAPPs). While offering robust reasoning capabilities over unstructured data, the implementation of SAPPs is often constrained by high financial costs and slow execution latencies. Palimpzest emerges as a declarative framework that decouples the logical specification of a program from its physical implementation, enabling automatic optimization. This internship study focuses on reproducing experiments on the Palimpzest system by integrating a novel cost-based optimizer, Abacus. Experiments were conducted by executing various standard workloads, including Legal Discovery and Real Estate Search, to compare the execution plans generated by Abacus against naive baseline implementations. Experimental results indicate that the integration of Abacus with new models is capable of identifying more efficient execution plans. This study validates the claims made in the original paper.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Large Language Models (LLM), Semantic Analysis Applications (SAPPs), Palimpzest, Abacus, Optimasi, Optimizations
Subjects:	Q Science Q Science > Q Science (General) Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA402 System analysis.
Divisions:	Faculty of Industrial Technology > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Mathias Adya Diwangkara Suganda
Date Deposited:	13 Jan 2026 01:26
Last Modified:	13 Jan 2026 01:26
URI:	http://repository.its.ac.id/id/eprint/129523

Actions (login required)

View Item