1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
//! Deterministic fake GPUs for UI development and demos (`--mock`).
use super::{GpuBackend, GpuProcess, GpuSnapshot, ProcKind};
use anyhow::Result;
/// Mock pids start well above anything the kernel hands out (Linux caps
/// pid_max at 2^22, and the classic default is 32768), so a demo can never
/// point a row at a live system process the way plain `1..3n` did.
const PID_BASE: u32 = 1_000_000;
/// Synthetic rows: mock pids resolve to nothing on this host, so supply the
/// host columns here instead of rendering a table full of "?".
const FAKE: [(&str, &str); 6] = [
("alice", "python train.py --epochs 30 --amp"),
("bob", "ollama runner --model llama3:70b"),
("alice", "blender -b scene.blend -f 120"),
("root", "/usr/lib/xorg/Xorg vt2 -displayfd 3"),
("bob", "ffmpeg -hwaccel cuda -i in.mkv out.mp4"),
("carol", "/usr/lib/firefox/firefox -contentproc"),
];
pub struct MockBackend {
tick: u64,
count: usize,
}
impl MockBackend {
pub fn new(count: usize) -> Self {
Self { tick: 0, count }
}
fn wave(&self, phase: f64, period: f64) -> f64 {
let t = self.tick as f64;
50.0 + 50.0 * (t / period * std::f64::consts::TAU + phase).sin()
}
}
impl GpuBackend for MockBackend {
fn name(&self) -> &'static str {
"mock"
}
fn poll(&mut self) -> Result<Vec<GpuSnapshot>> {
self.tick += 1;
// Test hook: GPUR_MOCK_FAIL=N fails every Nth poll to exercise the
// graceful-degradation path.
if let Some(n) = std::env::var("GPUR_MOCK_FAIL")
.ok()
.and_then(|v| v.parse::<u64>().ok())
&& n > 0
&& self.tick.is_multiple_of(n)
{
anyhow::bail!("simulated driver reset (tick {})", self.tick);
}
let total = 16 * 1024 * 1024 * 1024u64;
let gtt_total = 8 * 1024 * 1024 * 1024u64;
let gpus = (0..self.count)
.map(|i| {
let util = self.wave(i as f64 * 1.3, 37.0 + 11.0 * i as f64);
let vram = self.wave(i as f64 * 2.1 + 0.7, 97.0);
GpuSnapshot {
name: format!("Mock GPU {i}"),
// Synthetic, but stable across polls and across runs,
// which is all the identity contract asks for.
device_id: Some(format!("mock:{i}")),
integrated: i == 0,
utilization_pct: Some(util),
mem_util_pct: Some(util * 0.6),
video_util_pct: (i % 2 == 0).then_some(util * 0.4),
enc_util_pct: (i % 2 == 1).then_some(util * 0.3),
dec_util_pct: (i % 2 == 1).then_some(util * 0.5),
throttle: (util > 95.0).then(|| "power-limit".to_string()),
// GPU 0 is a unified-memory part: no local pool at all, so
// everything it spends is the system-backed one below.
// The demo has to carry that shape — it is what every
// iGPU, APU and Apple Silicon card looks like, and a
// fleet of identical dGPUs never exercises its rendering.
vram_used_bytes: (i > 0).then(|| (total as f64 * vram / 100.0) as u64),
vram_total_bytes: (i > 0).then_some(total),
temperature_c: Some(45.0 + util * 0.4),
power_w: Some(60.0 + util * 2.4),
power_limit_w: Some(320.0),
fan_pct: Some((util * 0.9).min(100.0)),
clock_mhz: Some(1200 + (util * 12.0) as u64),
mem_clock_mhz: Some(9000),
// GPU 1 demos the PCIe downgrade warning.
pcie_gen: if i == 1 { Some(3) } else { Some(4) },
pcie_width: if i == 1 { Some(8) } else { Some(16) },
pcie_max_gen: Some(4),
pcie_max_width: Some(16),
pcie_rx_kbs: Some((util * 8000.0) as u64),
pcie_tx_kbs: Some((util * 3000.0) as u64),
fan_rpm: Some(800 + (util * 14.0) as u64),
temp_junction_c: Some(50.0 + util * 0.45),
// Memory sensor sits between core and junction.
temp_mem_c: Some(48.0 + util * 0.42),
// NVML-style performance state: P0 under load, P8 at idle.
perf_level: Some(if util > 50.0 {
"P0".into()
} else {
"P8".into()
}),
// On GPU 0 this is the whole of its memory; on the others
// it is the spill pool beside their VRAM, so it moves
// with the fill level rather than sitting still.
gtt_used_bytes: Some(if i == 0 {
(gtt_total as f64 * vram / 100.0) as u64
} else {
(util * 2e7) as u64
}),
gtt_total_bytes: Some(gtt_total),
volt_mv: Some(700 + (util * 4.0) as u64),
}
})
.collect();
Ok(gpus)
}
fn driver_info(&self) -> Option<String> {
Some("mock driver 1.0".into())
}
fn processes(&mut self) -> Vec<GpuProcess> {
let util = self.wave(0.4, 53.0);
// A few rows per GPU so the table can overflow and scroll in demos.
(0..self.count * 3)
.map(|i| {
// Row 0 is gpur itself, left un-enriched so sysinfo fills in
// genuine host columns for one row; the rest are fabricated.
let own = i == 0;
let (user, command) = FAKE[i % FAKE.len()];
GpuProcess {
pid: if own {
std::process::id()
} else {
PID_BASE + i as u32
},
gpu_index: i % self.count,
kind: if i % 3 == 0 {
ProcKind::Graphics
} else {
ProcKind::Compute
},
gpu_util_pct: Some((util * (0.9 - 0.1 * i as f64)).max(0.0)),
// Fabricated, but fabricated as a figure: the demo rig is
// pretending to be a backend that can read memory.
gpu_mem_bytes: Some((3000 >> i.min(8)) as u64 * 1024 * 1024 + 64 * 1024 * 1024),
user: (!own).then(|| user.to_string()),
command: (!own).then(|| command.to_string()),
cpu_pct: (!own).then(|| (util * (0.5 - 0.05 * i as f64)).max(0.0) as f32),
host_mem_bytes: (!own).then(|| (400 >> i.min(6)) as u64 * 1024 * 1024),
container: None,
}
})
.collect()
}
/// Mock pids are fabricated; nothing here may be signalled.
fn can_signal(&self) -> bool {
false
}
}