Add C++ benchmarks for local IO throughput

[vuule]

Benchmarks should cover different workloads (size/number of reads/writes).
Multi-threaded benchmarks can also be considered.
Is possible, the throughput can be compared to the results from utilities like fio and gdsio.

[GregoryKimball]

@kingcrimsontianyu thanks for completing the first step in #659. What types of benchmarks would be most important to include before we complete the spirit of this issue?

[kingcrimsontianyu]

At minimum, the benchmarks should include the sequential read (write) and random read (write) I/O patterns for KvikIO's FileHandle::pread/pwrite, and report throughput (GiB/s) and IOPS (I/O per second) for these two patterns respectively.

Then we may extend the benchmark to mixed I/O pattern, latency measurement, MmapHandle::pread.

[JigaoLuo]

One point I’d like to raise, as I discovered this WIP today: I hope the tool could also support NVMe‑over‑Fabric.
In principle, it should not require extra work, since once NVMeoF is set up, both fio and gdsio run seamlessly on it, once gdscheck shows NVMeoF as well.

[kingcrimsontianyu]

Good reminder. I think the new per-drive thread pool feature is expected to continue to work (thought I don't have a machine with NVMe-OF to verify atm). The get_block_device_info will still resolve the file path to a specific device such as nvme0, which is now a remote target instead of a local block device.

[JigaoLuo]

I can help once the framework is ready, so that I can extend it to NVMe‑oF and test it.

That being said, the remote SSDs from NVMe‑oF still follow the NVMe namespace convention, that the function get_block_device_info must work as well. For example, in our setup, we have 10 PCIe-attached SSDs, and I mount 8 remote NVMe‑oF SSDs. (Please ignore the Usage columns, as it seems nvme‑cli is not functioning correctly for some unknown reason.)

(For NVMe‑oF with GDS, the process is quite involved and a pain: selecting GPU driver, installing OFED, then GDS, and finally setting up NVMe‑oF. My experience is mainly with vanilla NVMe‑over‑RDMA, and I don’t have knowledge of the other vendor implementations listed at the end of gdscheck)

[kingcrimsontianyu]

I can relate, as you are playing the role of a developer, a devops engineer and a technical writer. 😄
Is the GDS document clear enough to guide you through the intricate installation steps? If there is any certain part(s) in need of improvement, I can forward your concern or suggestion to the GDS team.