Three papers accepted to ASPLOS 2021
ASPLOS 2021 - April 19-23 2021
Congratulations to Boris, Mike and their respective teams for getting three papers accepted at ASPLOS 2021 between them.
ASPLOS is the premier forum for interdisciplinary systems research, intersecting computer architecture, hardware and emerging technologies, programming languages and compilers, operating systems, and networking.
Details of the successful papers can be found below:
1) Record-and-Prefetch Serverless Functions Orchestration.
Dmitrii Ustiugov, Plamen Petrov, Boris Grot, and collaborators from EPFL.
Serverless functions are an exceedingly popular way to deploy cloud services. Many functions are invoked infrequently, in which case they experience high start-up latencies. Our work provides a detailed characterization of function start-up latency in a state-of-the-art deployment. Using the insights from the characterization, the paper proposes a lightweight software scheme to rapidly restore the function’s core memory footprint, hence significantly lowering function start-up latency. We are also releasing the first open-source end-to-end serverless infrastructure integrating AWS Firecracker, Containerd and Kubernetes. Available now!
2) PTEMagnet: Fine-grained Physical Memory Reservation for Faster Page Walks in Public Clouds.
Artemiy Margaritov, Dmitrii Ustiugov, Amna Shahab, Boris Grot.
This work identifies address translation as an important performance bottleneck in virtualized cloud deployments and shows that the problem is caused by the Linux memory allocator’s fragmentation of memory, which reduces caching efficiency for page table entries. In response, the work proposes a simple approach to improve page table locality through minimal changes to the Linux kernel, thus improving performance of important cloud workloads.
3) Neural Architecture Search as Program Transformation Exploration Jack Turner, Elliot Crowley, Michael O'Boyle We express neural architecture operations as program transformations whose legality depends on a notion of representational capacity. This allows them to be combined with existing transformations into a unified optimization framework. This unification allows us to express existing NAS operations as combinations of simpler transformations, as well as exploring new ones. We prototyped the combined framework in TVM and were able to find optimizations across different DNNs that significantly reduce inference time.