On October 12, Meta hosted its first in-person Networking and Communications Faculty Summit since 2019. Meta engineers across Network Infra teams, along with 20 academic researchers and PhD students from leading research universities in the networking space, met in New York City to discuss warehouse-scale AI networking, deep learning workloads at the network interface, data center routing verification, and network traffic and ML for network planning, among other topics.

“Collaboration with the research community has always been an important part of how we advance the state of the art in networking and communication at Meta,” says James Zeng, Senior Engineering Manager in Network Infra. James kicked off the summit, presenting an agenda that featured internal and external speakers, including research award winners from last year’s call for proposals and Meta Research PhD Fellows.

“This year we were very excited to welcome our research award winners and PhD Fellows in person at our New York campus,” he added, “and to discuss the new advancements and research problems the community is working on, and how these areas will transform in the future.”

The summit featured talks from six Meta researchers: Adi Gangidi, Mohab Gawish, Varun Gupta, Petr Lapukhov, Brandon Schlinker, and Yimeng Zhao. The discussions offered an opportunity to exchange views with academics and present Meta’s own research.

Research award winners and PhD Fellows who presented included Mohamed Abdelfattah (Cornell University), Gianni Antichi (Queen Mary University of London), Claudio Cicconetti (National Research Council), Soudeh Ghorbani (Johns Hopkins University), and Stewart Grant (University of California San Diego).

The summit concluded with a panel discussion with Meta Network Infra team members Adi Gangidi, Mohab Gawish, Varun Gupta, Petr Lapukhov, Brandon Schlinker, Srikanth Sundaresan, and James Zeng.

The summit wrapped up a busy year of research for our Network Infra group. Throughout 2022, Meta researchers published in several international academic conferences, including NSDI, Hot Interconnects, SIGCOMM, and IMC, with the following papers:

• Check-N-Run: A checkpointing system for training deep learning recommendation models
• Evolvable network telemetry at Facebook
• NetHint: White-box networking for multi-tenant data centers
• Network entitlement: Contract-based network sharing with agility and SLO guarantees
• Flash: Fast, consistent data plane verification for large-scale network settings
• A microscopic view of bursts, buffer contention, and loss in data centers
• Cross-layer diagnosis of optical backbone failures
• Impact of RoCE congestion control policies on distributed training of DNNs