GSoC'24 Final Report

[fr0m-scratch]

Google Summer of Code 2024 Final Report

Project: Live Patching Distributed ClickHouse Cluster with bpftime

Mentor: victoryang00
Student: fr0m-scratch
Source Code: https://github.com/eunomia-bpf/Clickhouse-Uring/tree/main

---

Overview

The goal of this project was to enhance the performance of distributed ClickHouse by leveraging io_uring and eBPF. bpftime and USDT were used to achieve minimal modification to ClickHouse source code and provide an interface for further optimization with in-kernel storage function resubmission to NVMe driver. A second part of this project was to achieve in-kernel raft server implementation with XDP(eXpress Data Path); but unfortunately this part is largely undone, and remains in the POC phase of simply dispatching heartbeat packet at the XDP edge.

Objectives and Deliverables

1. io_uring supported server and client prototype
2. More io_uring helper for bpftime
3. io_uring batched read & write POC with bpftime
4. A benchmark ClickHouse cluster with two shards and three keepers
5. ClickHouse source code patches of userspace override interface for pread and double buffering prefetch using USDT probes
6. Proof of concepts with batched pread using io_uring and userspace override interface in No.4
7. Proof of concepts of dispatching raft heartbeat packets with XDP

Future Work

1. The No.5 in the deliverables was tested only with io_uring batch in userspace code but not in bpf programs with io_uring helpers provided by bpftime because of unfixed bugs with USDT compatibility. This will be tested when the bpftime USDT support is fixed.
2. Implement the in-kernel raft server for ClickHouse, which is undone from the original plan.
3. Provide complete benchmark and documentation for all artifacts above.

The remaining work of this project above is expected to be completed by the end of the year. After that, I plan to continue contribute to the eunomia-BPF community and delve deep into the actual internals of userspace BPF—something regretfully I didn't have time for this summer.

Conclusion

Participating in GSoC 2024 was an enriching experience. I am grateful for the guidance and support I received from my mentor and the eunomia-BPF community.

Appreciate all,
Jamal Cao