OSDI’25

My random murmur about OSDI’25

A couple of things I was looking forward to before this exciting trip…

First off, I was eager to chat with folks. I wasn’t presenting our paper, but it never hurts to spread the word to a broader audience. My buddy Shreesha did a fantastic job presenting our work, Speclog. Speclog is a new take on the shared log that optimizes for end-to-end latency. The shared log is a powerful abstraction—it gives a consistent view of a log in distributed systems and is widely used across many distributed applications. Speclog brings speculative execution into the picture to cut down end-to-end latency, which is crucial for performance in a huge number of use cases. Overall, I had some really satisfying conversations with folks. As an early-stage grad student, I find the poster session to be the easiest and most natural way to strike up a chat. I just wish the organizers gave us more time for posters—they always feel too short.

The talks, on the other hand, were a bit underwhelming. For a conference as prestigious as OSDI, I expected presenters to be ultra-prepared, slides polished, delivery sharp, and ready to handle any tough questions. That didn’t always seem to be the case. Some talks had confusing slides or awkward pacing, which honestly made it hard to stay engaged or feel motivated to ask questions. The Q&A sessions were quieter than I expected, and in a few cases, presenters didn’t seem fully prepped and gave pretty weak answers.

That said, one bright spot was how OSDI didn’t go overboard on the ML content. MLSys was capped at just three sessions—the rest were pure systems. I’ve always felt systems is the bedrock of computation. Web, big data, AI—whatever it is, systems quietly powers it all behind the scenes. These days, I don’t think we can (or should) completely ignore ML, but I do think the systems community needs to hold on to problems that are pure and deeply rooted in our field. OSDI did a great job providing space for that.

A couple of papers stood out to me. One was Tigon—to my knowledge, they’re the first to tap into CXL shared memory to speed up existing applications. Very cool. Another was Mako. They also use speculative execution, but in a totally different context. It’s a great example of how our community loves seeing long-standing ideas applied in clever, fresh ways. I can think of a dozen other great papers that follow this pattern.

Believe it or not, this was my first time visiting Boston—even after nearly 8 years in the U.S.! It’s always tricky to squeeze in sightseeing during a conference, but I managed to sneak in a quick walk around the MIT quad and CSAIL. We even tried sneaking into the upper levels of CSAIL to find PDOS, but failed. Can someone tell me—is PDOS an actual office, or is it just a conceptual label? Either way, MIT looked awesome. Boston itself was also super pleasant—appropriately sized (I grew up in Shanghai, so I have some low-key metropolitan-phobia), and I loved how close downtown is to the airport.

I came into this OSDI with a light mind—no presentation duties, and we accidentally messed up poster registration, so no poster from us either. It was chill not having responsibilities, but I’m hoping the next conference I attend, I’ll be presenting something of my own.