Title: Low Latency Operating System
Technical Area: System Software
While we are transforming a legacy large-scale complex system so that it can run on serverless platform. The Linux kernel has become a difficult technical bottleneck for us. When the system load is high, Linux is difficult to guarantee a certain low latency. Scheduling latency, IO latency, network latency and memory allocation latency are related to workload and how busy the system is. On the serverless platform, we can't get low latency and high throughput at the same time. So, we should have a new operating system that has creative processes scheduler, memory management, network and IO stack. With the new OS, a process can be started up below a millisecond, hardware resources can be made full use of. Meanwhile, the various latency in the operating system are stable and predictable. Based on this new operating system, we can achieve a serverless platform with maximum efficiency.
We hope a proof of concept for the low latency operating system can be implemented. We can assume that the operating system runs on the various latest hardware, such as Apache Pass, RDMA, and so on. We can also implement some core logic with FPGA chip. The PoC may not meet the standards of a production system. But we want it to have the following characteristics:
- The process can be started up below one millisecond.
- Its scheduling latency can be stable at the microsecond level.
- Its network and IO latency can be stable at the microsecond level.
- It has a stable memory management system without any jitter.
- When the system runs some typical workloads (such as Database, memory cache or webserver) with very high loads, the operating system cost (OS Tax) should be less than 2%.
- When the CPU utilization of the system reaches 80%, the application respond time increases by no more than 10%.
Related Research Topics
- Process scheduler in user space
- Micro Kernel
- Hardware and software combination technology.