Projects
What I am working on now, and systems I have built. Papers are listed under publications.
Current
Cost-efficient LLM inference
Software-systems-level optimization for LLM inference, aimed at the cost break-even point: where self-hosting open-weight models becomes cheaper than paying for commercial API access.
In-network computing over quantized data
Hardware support so that in-network computation can operate directly on quantized data, accelerating distributed training and inference.
Datacenter congestion control
Congestion control for high-performance datacenter networks, working with DPDK and RDMA.
Past
P4TVal: baseline and evaluation
P4TVal detects unauthorized changes to the forwarding rules in a programmable switch. My part was the comparative evaluation:
- Implemented REV (Zhang et al., IEEE/ACM ToN 2020), the prior state of the art, in P4 on BMv2 and Mininet, as the baseline P4TVal is measured against.
- Compared attack-detection time as the network grows from 2 to 32 switches. P4TVal detects table-rule manipulation 10× faster, and its detection time stays constant while REV's grows with hop count.
- Measured P4TVal's data-plane overhead on an Aurora 610 Intel Tofino switch: about 50 ns added per packet, over 5,000 packets.
P4Auth live demonstration
Built the live demo for P4Auth, a system that secures in-network traffic control by authenticating the messages it relies on, between controller and switch and between switches. The demo runs HULA load balancing in three configurations side by side: normal operation, a compromised switch, and the compromised switch with P4Auth in place. Acknowledged in the paper.