duclos.dev     about  posts  resume  contact


Master Thesis — VM Selection for Financial Exchanges
________________________________________________________________________________

2024-10-22 · tags: research, cloud

After a challenging year I finally defended my master thesis on "VM Selection
for Financial Exchanges in the Cloud", my final credit at the Technical
University of Munich. Co-advised by PhD. Muhammad Haseeb and
Dr. Anirudh Sivaraman (Systems@NYU), and PhD. Navidreza Asadi and
Prof. Dr.-Ing. Wolfgang Kellerer (LKN, TUM). Code and docs:
https://github.com/duclos-cavalcanti/TreeBuilder



[001] Abstract
________________________________________________________________________________

Cloud computing has revolutionized the deployment of software services by
providing virtualized resources rented on-demand. Financial exchanges are
exploring migrating to the cloud for scalability and reduced cost. But the
cloud's inherent network variance and lack of native multicast challenge the
fair, low-latency dissemination modern exchanges require.

Jasper implements a multicast solution for cloud-based exchanges, reaching
ultra-low latency via an overlay proxy tree, kernel bypass, and clock
synchronization. It does not fully address "stragglers" — underperforming VMs
that degrade system performance.

TreeBuilder enhances Jasper by intelligently selecting nodes from a pool to form
a multicast tree, specifically targeting straggler VMs. Using straightforward
heuristics (WEIGHTED, P90, P50) it improves latency while staying lightweight
and modular. An instance of TreeBuilder's BEST-WEIGHTED tree showed the lowest
90th-percentile latency for its worst receiver at 312 us — ~7.7% below
LemonDrop's 338 us — while remaining an accessible, extensible technique with
considerable scalability.

________________________________________________________________________________

Daniel Duclos-Cavalcanti (C) 2026 · built with make + sed · no JavaScript