No Provisioned Concurrency: Fast RDMA-codesigned Remote Fork for Serverless Computing
summary
MITOSIS implements an OS primitive called remote fork that co-designs Linux kernel behavior with RDMA to let a machine read parent pages on demand and avoid checkpointing or full state transfer. It launches a container in ~5 ms cold, forks over 10,000 containers across machines within one second, and reduces function tail latency by 89% while cutting memory usage versus cached instances. Integrated with FN, it enables no provisioned concurrency and accelerates serverless state transfer, improving workflow runtime by 86%.
AI-written summary · about
chapters (10)
- 0:00Mitosis and remote Fork
- 1:28Slow container starts
- 3:14No provisioned concurrency
- 5:46Existing Fork limits
- 6:52Remote Fork challenges
- 8:17RDMA co-design for Fork
- 10:48Mitosis benefits
- 11:32Auto scaling with Mitosis
- 13:19Fast state transfer
- 14:45Implementation and conclusion
related talks
17mTowards High-Performance Transactional Stateful Serverless Workflows with...
33mFission: Serverless Functions for Kubernetes
16mFork in the Road: Reflections and Optimizations for Cold Start Latency in Production...
30mOrion: A Distributed File System for Non-Volatile Main Memory and RDMA-Capable Networks
19mA Secure, Fast, and Resource-Efficient Serverless Platform with Function REWIND
21mFine-Grained Isolation for Scalable, Dynamic, Multi-tenant Edge Clouds
17mServerless Functions Made Confidential and Efficient with Split Containers
36mFission: a Serverless Function Framework on Kubernetes [B]
