We designed Source and Intermediate Representation (IR) containers. In IR containers, we implement a pipeline where containerized applications are compiled to LLVM IR, and adapted to the target architecture (e.g., vectorization) at deployment time, ensuring peak performance.

Traditional containers lock in compilation decisions too early, hurting performance on diverse HPC systems. We delay performance-critical decisions until the target system is known, avoiding the traditional container trade-off between portability and hardware-specific optimizations.

This year at #SC25, we will present XaaS containers, the Best Paper Finalist at 🏆! We propose a new paradigm for performance portability to bridge the gap between #cloud productivity and #HPC performance 🚀

#HPCignites @thoefler.bsky.social

Our evaluation shows that C++ serverless functions can scale to 512 parallel workers with a double-digit millisecond overhead. Using the example of ray tracing, we demonstrate a speedup of up to 59x. From a 60-second execution time to 1 second with hundreds of functions and a minimal cost increase.

With Cppless, C++ is no longer a second-class citizen in serverless. We introduce alternative entry points and implement lambda reflection, making C++ functions as easy to deploy as Python.

Cold starts? Just 10-12 milliseconds and up to 60x faster than Python functions!

🚀Cppless enables single-source #serverless programming in C++! No more splitting functions from the main code - just add a few lines to accelerate your program by offloading to systems like AWS Lambda. Our paper on Clang-based #FaaS for C++ has just been accepted at ACM TACO.

Reasoning Language Models: A Blueprint

📜 arxiv.org/abs/2501.11223
x1 on GitHub: github.com/spcl/x1

We introduce x1, a modular implementation for rapid RLM prototyping, provide key insights (multi-phase training for policy/value models, importance of familiar training distributions) and discuss scalable RLM cloud deployments and how RLMs fit into the broader LLM ecosystem.

We propose a blueprint that organizes RLM components into a modular framework, based on an analysis of all RLM works, which incorporates reasoning structures (chains, trees, graphs) and strategies (e.g., #MCTS), Reinforcement Learning and supervision schemes.

Reasoning language models (#RLMs), such as o1 or QwQ, have redefined AI's problem-solving capabilities by extending #LLMs with advanced reasoning mechanisms. Yet, their high costs, proprietary nature, and complex architectures present accessibility and scalability challenges.

Kyle Chard opens panelists statements at #SC24 panel on #serverless #HPC. Join us in room B314!

The #SC24 panel on #serverless #HPC happens today at 10:30 am! Looking forward to the discussion on the future Cloud & HPC with our panelists: @ianfoster42.bsky.social, @thoefler.bsky.social, Satoshi Matsuoka, Kyle Chard, and Devesh Tiwari!

Great attendance at the #SC24 panel on performance portable containers #XaaS! Interesting conclusions on the future of #HPC and #cloud, in a discussion between @danernst.bsky.social, @ianfoster42.bsky.social, Torsten Hoefler, and Thomas Schulthess!

Is #serverless a fad that will disappear in a few years, or will it accelerate #HPC - #Cloud convergence and change the #Supercomputing landscape?

Join us at the #SC24 panel "Serverless HPC: Challenges, Opportunities, and Future Prospects for Accelerated Cloud Computing"!

Can high-performance containers offer performance portability and unified abstraction for the converged #HPC and #Cloud? Join us at the #SC24 CANOPIE-HPC workshop for a panel on the changing HPC landscape.

📅 Sunday, 17th November
⏰ 11:40 AM EST
🏢 #Supercomputing Room B313