add hardware topology module

[bg-furiosa]

PR Description

• add a hardware topology module to the device API that can detect the server's hardware topology and measure the distance between devices.
• this is back porting of go implementation[1][2].

What type of PR is this?

• /kind feature

Special notes for reviewer

./show_topology
Authorization required, but no authorization protocol specified
+--------+----------------------+----------------------+----------------------+----------------------+
| Device | npu0                 | npu1                 | npu2                 | npu3                 |
+--------+----------------------+----------------------+----------------------+----------------------+
| npu0   | LinkTypeSoc          | LinkTypeCPU          | LinkTypeInterconnect | LinkTypeInterconnect |
+--------+----------------------+----------------------+----------------------+----------------------+
| npu1   | LinkTypeCPU          | LinkTypeSoc          | LinkTypeInterconnect | LinkTypeInterconnect |
+--------+----------------------+----------------------+----------------------+----------------------+
| npu2   | LinkTypeInterconnect | LinkTypeInterconnect | LinkTypeSoc          | LinkTypeCPU          |
+--------+----------------------+----------------------+----------------------+----------------------+
| npu3   | LinkTypeInterconnect | LinkTypeInterconnect | LinkTypeCPU          | LinkTypeSoc          |
+--------+----------------------+----------------------+----------------------+----------------------+

[libc-furiosa]

Before reviewing at the code level, I want to comment on the design.

First, we need to simplify the implementation of Topology.
Right now, the struct is behaving mutable with state, but it can actually have deterministic values at initialization time.
This makes the interface more compact and clear.

Second, if you need to add a dependency on hwloc, it would be better to build it from source code and include it in the build output via a static link rather than relying on some arbitrary OS package manager.
Meanwhile, there are hwloc implementations on crates.io. Please comment if you think we should use them.
(e.g., https://github.com/HadrienG2/hwlocality)

[libc-furiosa]

I started by looking at the big flows. I'm not completely done with the review, but I'll continue with the review after seeing the changes for comments.

[libc-furiosa]

I left some comments. Let's just resolve those and be done with it.

[libc-furiosa]

This can be omitted.

[libc-furiosa]

What is this dependency? If it is not used, you can remove this line and remove the libc dependency as well.

[libc-furiosa]

Suggested change

	use furiosa_device::{list_devices, topology, DeviceError};
	use furiosa_device::{list_devices, topology::Topology, DeviceError};

[libc-furiosa]

Suggested change

	let topology = topology::Topology::new(devices.clone())?;
	let topology = Topology::new(devices.clone())?;

[libc-furiosa]

Suggested change

	let mut rows = vec![];
	let mut header = vec!["Device".cell().bold(true)];
	let mut rows = Vec::with_capacity(devices.len() + 1);
	let mut header = Vec::with_capacity(devices.len() + 1);
	header.push("Device".cell().bold(true));

[libc-furiosa]

Suggested change

	let mut link_type: LinkType = LinkTypeUnknown;
	if dev1_bdf == dev2_bdf {
	link_type = LinkTypeSoc
	} else {
	link_type = topology_provider.get_common_ancestor_obj(&dev1_bdf, &dev2_bdf)?;
	}
	let link_type = if dev1_bdf == dev2_bdf {
	LinkTypeSoc
	} else {
	topology_provider.get_common_ancestor_obj(&dev1_bdf, &dev2_bdf)?
	}

[libc-furiosa]

Suggested change

	topology_matrix.insert(key, link_type);
	topology_matrix.entry(key).or_insert(link_type);

[libc-furiosa]

Is there a reason why the access specifier for this function is pub? Are there other places it's used?

[libc-furiosa]

Is there a reason why the access specifier for this function is pub? (2)

[libc-furiosa]

Let's make sure this is the dependency we actually need.

[libc-furiosa]

LGTM 👍