# Systems Fun at HotOS DevFeed: [Systems Fun at HotOS]() Original publisher: [Read original article]() Author: Marc Brooker Published: 2025-06-02T00:00:00Z Content type: article Language: en Sources: [Marc Brooker's Blog](), [Marc Brooker's Blog]() Topics: [systems](), [Concurrency](), [Serverless](), [Latency](), [Cloud](), [Cache]() Tags: [batching](), [cache](), [cloud](), [concurrency](), [latency](), [performance](), [serverless](), [systems]() ## AI overview The article reflects on attending HotOS for the first time and highlights several papers and talks. It discusses integrating NIC scheduling with OS state, the relevance of this approach to cloud and Serverless workloads, batching's context-dependent effects on latency and throughput, and a proposed dynamic replacement for fork using posix_spawn. The supplied text ends mid-discussion of the final paper. ## Source excerpt Systems Fun at HotOS One day somebody will tell me what systems means. Last week I attended HotOS1 for the first time. It was super fun. Just the kind of conference I like: single-track, a mix of academic and industry, a mix of normal practical ideas and less-normal less-practical big thinking. I went partially because a colleague twisted my arm, and partially because of this line in the CFP: The program committee will explicitly favor papers likely to stimulate reflection and discussion. That sounds like a good time. Some Papers or Talks I Enjoyed I thought I'd give a shout-out to some of the papers I enjoyed the most. These aren't the best papers, just ones I particularly liked for my own arbitrary reasons2. The NIC should be part of the OS by Xu and Roscoe exposes a lot of what was traditionally the OS kernel's state (like the run queue for each core) to the NIC, allowing the NIC to use that state to decide which packets to deliver and when. Their goal is to do better than existing kernel bypass methods. After all, many modern cloud applications are short bursts of compute between waiting for packets. They do this with CXL 3.0's cache-coherent peripheral interconnect, but I don't think peripheral cache coherence is actually critical to the big picture here. This approach of pushing scheduling work down to the NIC seems especially interesting for straight packet processing workloads, and for Serverless workloads, both of which tend to be very run and wait heavy with high concurrency. Just super fun deep systems work3. Batching with End-to-End Performance Estimation by Borisov et al. This paper starts by questioning the common wisdom that batching is good for throughput and bad for latency, by presenting a set of scenarios that show that it can be good or bad for both latency and throughput depending on small timing differences. I love questioning the common wisdom, so this is my jam. They then point out that this situation can be improved by making batching logic