Alireza Farshin

We show that FAJITA can process up to 170Mpps with an NF consisting of a flow counter +a rate limiter and a load balancer. We then show how to improve by 50% the throughput of NF chains that match on different fields of the header, processing up to 90Mpps.

We address these challenges by devising a careful NF packet processor (on top of FastClick), which we call FAJITA, that carefully orchestrate the prefetching and disaggregation of any information that must be retrieved by any core to process a packet.

We discovered that: - Chasing a single pointer on a CPU may cost you 10 Mpps - Many NF chains retrieve flow states based on different packet header fields, breaking the "exclusive memory access" guarantees of high-speed multi-core CPUs In our CoNEXT'24 paper... (see comments)

Do large language models facilitate network configuration? We investigate opportunities and challenges in operating network systems using LLMs in our accepted ACM CoNEXT'24 paper. Thanks, @WangChangjie13 @Skazza_94 @alirezafarshin @si_ferlin @dmkostic. More info in the comments.

I would appreciate your help in spreading the word about the extended deadline (May 13) for our Assoc. Prof. in Systems job opening: kth.se/en/om/work-at-kth/led… Please do not hesitate to contact Dejan Kostic if you have any questions.

Do you want to know more about SmartNIC (CPU-based) performance? Come to see what Georgios and @TomBarbette discovered at their PAM talk on Mar 30, 7pm CEST! One spoiler: multi-table packet processing and rule insertions are expensive! Joint work with D. Kostic and G. Maguire