George Karlos

Research

My interests sit at the intersection of HPC/AI, networking, and programming languages. I am currently working on in-network computing.

Some of the technologies I'm interested in (in no particular order):

• Programmable switches
• GPUs
• DPUs
• C / C++
• LLVM
• P4
• CUDA
• MPI
• DPDK
• RDMA

Small cluster with programmable networking equipment that I maintain and use for most of my experiments.

Available hardware

• 6x nodes with AMD EPYC 7443P 24-C, 256GB RAM
• 4x Nodes with 2x Intel Xeon E5-2640 v3 8-C, 64G RAM
• 1x Edgecore DCS810 (Tofino2) 32x400G Programmable switch
• 1x Netberg Aurora 710 (Tofino1) 32x100G Programmable switch
• 2x NVIDIA A30 GPUs, 6x NVIDIA P100 GPUs, 6x NVIDIA A2 GPUs
• 10x NVIDIA Connectx-6 Dx EN 2x 100G
• 2x NVIDIA Bluefield3 400G DPUs
• 1x Xilinx Alveo SN1022 2x 100G SmartNIC

Talks

• Supercomputing'24 - NetCL: A Unified Programming Framework for In-network Computing
• Hotnets'21 - Don't you worry 'bout a packet: Unified Programming for In-network Computing
• Netys'21 - Fault-tolerant termination detection with Safra's algorithm
• Memory Access Monitoring of GPU kernels (MSc thesis)

Notes

• ML training parallelism strategies
• Notebook LM — NetCL

Contact

I am generally reachable at g.karlos AT {vu.nl, upb.de}

You are particularly encouraged to contact me if you are a student looking for a project, or if you work on programmable dataplanes and think there is room for collaboration.