NVIDIA DCGM simulation

Simulator side of the GPU story. This page covers the simulator: the metric OID types, the GPU cycler extension, device profiles, resource file layout, and integration points. The OIDs a collector polls are in GPU simulation → Collector OID contract.

Overview

nl6 simulates NVIDIA GPU servers (DGX A100, DGX H100, HGX H200) with eight GPUs each. The metric vocabulary and the nvidia-smi / dcgmi command output are modelled on NVIDIA Data Center GPU Manager (DCGM). The simulation exposes GPU metrics via all three existing protocols: SNMP (per-GPU objects under NVIDIA's PEN, in a layout nl6 defines), SSH (nvidia-smi and dcgmi CLI), and REST API (DCGM-shaped HTTP endpoints).

What DCGM Exposes in the Real World

NVIDIA DCGM provides:

GPU utilization (compute and memory controller %)
GPU memory (used/total VRAM per GPU)
GPU temperature (current, shutdown threshold, slowdown threshold)
Power draw (current watts, power limit)
Fan speed (% of max)
GPU clock speeds (SM clock, memory clock in MHz)
PCIe throughput (TX/RX MB/s)
ECC error counts (single-bit corrected, double-bit uncorrected)
NVLink bandwidth (active links, throughput)
GPU health status (healthy, warning, critical)
Process/job info (running PIDs, memory per process)

Real DCGM exposes these through its API and the DCGM exporter's Prometheus /metrics endpoint, not through SNMP. nl6 exposes them via:

SNMP: OIDs under 1.3.6.1.4.1.5703, NVIDIA Corporation's IANA-registered Private Enterprise Number. The layout below the PEN is nl6's own (see below).
SSH: nvidia-smi command variants (nvidia-smi, nvidia-smi -q -d MEMORY|UTILIZATION|TEMPERATURE|POWER, nvidia-smi topo -m) and dcgmi discovery -l, dcgmi health -c
REST API: DCGM-shaped HTTP endpoints (/api/v1/gpu/status, /api/v1/gpu/devices, /api/v1/dcgm/health, etc.)

The objects below the PEN are nl6's own invention. 1.3.6.1.4.1.5703 is NVIDIA Corporation's real PEN, so sysObjectID correctly identifies a simulated DGX as an NVIDIA system. Nothing below the PEN is published by NVIDIA. NVIDIA ships no SNMP GPU MIB at all. Its own GPU telemetry story is DCGM and Prometheus, not SNMP. So 1.3.6.1.4.1.5703.1.1.1.* names no registered object and cannot be resolved against a real MIB. Treat the layout as nl6's own contract with its collectors, never as a vendor specification.

Migration. This arc was previously 1.3.6.1.4.1.53246, which the IANA enterprise-numbers registry allocates to Mailteck, S.A., a company unrelated to NVIDIA. A collector doing vendor detection resolved a simulated DGX as Mailteck. Every sub-identifier below the PEN was preserved exactly, so a downstream pollaris rule, oid_base, or vendor-detection prefix needs only 53246 replaced with 5703. No object was added, removed, or renumbered, and no OID's value or type changed.

Two things did change, and both are the point of the exercise rather than side effects. Every OID name under the arc moved. So did the three per-model sysObjectID values, from …53246.1.2.{1,2,3} to …5703.1.2.{1,2,3}, which is the string a collector reads to identify the vendor. The suffixes stayed distinct and in the same order, so the three models are still told apart.

The old arc is no longer served at all. An SNMPv2c or v3 GET under 1.3.6.1.4.1.53246 answers the RFC 3416 noSuchObject exception. An SNMPv1 manager gets error-status = noSuchName with the requested names echoed instead, since v1 has no exceptions (RFC 3584 §4.2.2.2.1). Either way the response carries no value, so an unmigrated rule collects nothing rather than collecting stale data.

Scope

Device Flavors (3 resource profiles)

Resource File	GPU Model	GPUs	VRAM/GPU	Total VRAM	System RAM
`nvidia_dgx_a100.json`	A100 80GB SXM	8	80 GB	640 GB	1 TB
`nvidia_dgx_h100.json`	H100 80GB SXM	8	80 GB	640 GB	2 TB
`nvidia_hgx_h200.json`	H200 141GB SXM	8	141 GB	1128 GB	2 TB

Each flavor gets its own resource directory under resources/ with SNMP, SSH, and API JSON data.

GPU-Specific Metric Types (new)

Extend MetricsCycler to generate cycling values for GPU metrics alongside existing CPU/memory/temperature metrics.

Phase 1: GPU Metric Types and Cycler Extension

1.1 New Metric OID Types (`metrics_oids.go`)

Add new MetricOIDType constants:

MetricGPUUtil       MetricOIDType = iota + 10  // GPU compute utilization %
MetricGPUMemUsed                                // GPU memory used (MB)
MetricGPUMemTotal                               // GPU memory total (MB, constant)
MetricGPUTemp                                   // GPU temperature (Celsius)
MetricGPUPower                                  // GPU power draw (Watts)
MetricGPUFanSpeed                               // GPU fan speed %
MetricGPUClockSM                                // GPU SM clock (MHz)
MetricGPUClockMem                               // GPU memory clock (MHz)

Add vendorOIDs entries for each NVIDIA resource file, using NVIDIA's enterprise OID prefix 1.3.6.1.4.1.5703. Each GPU (0-7) gets its own OID suffix. Example for GPU 0:

3.6.1.4.1.5703.1.1.1.1.5.0  → MetricGPUUtil       (GPU 0 utilization)
3.6.1.4.1.5703.1.1.1.1.6.0  → MetricGPUMemUsed    (GPU 0 memory used)
3.6.1.4.1.5703.1.1.1.1.7.0  → MetricGPUMemTotal   (GPU 0 memory total)
3.6.1.4.1.5703.1.1.1.1.8.0  → MetricGPUTemp       (GPU 0 temperature)
3.6.1.4.1.5703.1.1.1.1.9.0  → MetricGPUPower      (GPU 0 power draw)
3.6.1.4.1.5703.1.1.1.1.10.0 → MetricGPUFanSpeed   (GPU 0 fan speed)
3.6.1.4.1.5703.1.1.1.1.11.0 → MetricGPUClockSM    (GPU 0 SM clock)
3.6.1.4.1.5703.1.1.1.1.12.0 → MetricGPUClockMem   (GPU 0 memory clock)
...
3.6.1.4.1.5703.1.1.1.1.5.7  → MetricGPUUtil       (GPU 7 utilization)

That's 8 metric OIDs × 8 GPUs = 64 dynamic metric OIDs per device, plus the existing CPU/memory/temperature host metrics.

1.2 GPU Device Profile (`device_profiles.go`)

Add three new profiles:

var profileGPUServerA100 = DeviceProfile{
    CPUBaseMin: 20, CPUBaseMax: 55, CPUSpike: 25,
    MemTotalKB: 1024 * 1024 * 1024,  // 1 TB system RAM
    MemBaseMin: 40, MemBaseMax: 75, MemVariance: 12,
    TempBaseMin: 32, TempBaseMax: 48, TempSpike: 6,
}

var profileGPUServerH100 = DeviceProfile{
    CPUBaseMin: 25, CPUBaseMax: 60, CPUSpike: 28,
    MemTotalKB: 2 * 1024 * 1024 * 1024,  // 2 TB system RAM
    MemBaseMin: 45, MemBaseMax: 80, MemVariance: 12,
    TempBaseMin: 34, TempBaseMax: 50, TempSpike: 7,
}

var profileGPUServerH200 = DeviceProfile{
    CPUBaseMin: 25, CPUBaseMax: 60, CPUSpike: 28,
    MemTotalKB: 2 * 1024 * 1024 * 1024,  // 2 TB system RAM
    MemBaseMin: 45, MemBaseMax: 80, MemVariance: 12,
    TempBaseMin: 34, TempBaseMax: 50, TempSpike: 7,
}

Add to deviceProfileMap:

"nvidia_dgx_a100.json": profileGPUServerA100,
"nvidia_dgx_h100.json": profileGPUServerH100,
"nvidia_hgx_h200.json": profileGPUServerH200,

1.3 GPU Metrics in MetricsCycler (`metrics_cycler.go`)

Add a new GPUMetrics struct with per-GPU cycling arrays:

type GPUMetrics struct {
    gpuUtil     [numDataPoints]int    // GPU utilization % (0-100)
    gpuMemUsed  [numDataPoints]int64  // GPU memory used (MB)
    gpuMemTotal int64                 // GPU VRAM total (MB, constant)
    gpuTemp     [numDataPoints]int    // GPU temperature (Celsius)
    gpuPower    [numDataPoints]int    // GPU power draw (Watts)
    gpuFanSpeed [numDataPoints]int    // GPU fan speed % (0-100)
    gpuClockSM  [numDataPoints]int    // SM clock MHz
    gpuClockMem [numDataPoints]int    // Memory clock MHz
    utilIndex   uint32
    memIndex    uint32
    tempIndex   uint32
    powerIndex  uint32
    fanIndex    uint32
    clockIndex  uint32
}

Add gpuMetrics [8]*GPUMetrics field to MetricsCycler. These are only allocated when the device profile is a GPU server (checked by resource file name).

Each GPU gets its own seed (baseSeed + gpuIndex) so GPUs within the same server have different but correlated curves (e.g., GPU 0 might run hotter when GPU 1 is under heavy compute load).

GPU-specific profile parameters (new struct):

type GPUProfile struct {
    GPUCount       int   // Number of GPUs (8 for DGX/HGX)
    VRAMPerGPUMB   int64 // VRAM per GPU in MB
    GPUUtilMin     int   // Min GPU utilization %
    GPUUtilMax     int   // Max GPU utilization %
    GPUUtilSpike   int   // Utilization spike amplitude
    GPUTempMin     int   // Min GPU temp (Celsius)
    GPUTempMax     int   // Max GPU temp (Celsius)
    GPUTempSpike   int   // Temp spike amplitude
    GPUPowerMin    int   // Min power draw (Watts)
    GPUPowerMax    int   // Max power draw (Watts)
    GPUPowerSpike  int   // Power spike amplitude
    GPUClockSMBase int   // Base SM clock (MHz)
    GPUClockMemBase int  // Base memory clock (MHz)
}

Parameter	A100	H100	H200
VRAMPerGPUMB	81920	81920	144384
GPUUtilMin/Max	15-85	20-90	20-90
GPUTempMin/Max	35-75	35-80	35-78
GPUPowerMin/Max	100-400	150-700	150-700
GPUClockSMBase	1410	1980	1980
GPUClockMemBase	1512	2619	2619

1.4 Getter Methods for GPU Metrics

Add methods to MetricsCycler:

func (c *MetricsCycler) GetGPUUtil(gpuIndex int) string
func (c *MetricsCycler) GetGPUMemUsed(gpuIndex int) string
func (c *MetricsCycler) GetGPUMemTotal(gpuIndex int) string
func (c *MetricsCycler) GetGPUTemp(gpuIndex int) string
func (c *MetricsCycler) GetGPUPower(gpuIndex int) string
func (c *MetricsCycler) GetGPUFanSpeed(gpuIndex int) string
func (c *MetricsCycler) GetGPUClockSM(gpuIndex int) string
func (c *MetricsCycler) GetGPUClockMem(gpuIndex int) string

1.5 SNMP Handler Extension (`snmp_handlers.go`)

Update getMetricValue() to handle the new GPU metric types. The OID encodes the GPU index in the last component (e.g., .0 for GPU 0, .7 for GPU 7), so the handler must extract the GPU index from the OID suffix and call the appropriate getter.

Add a helper to parse GPU index from OID:

func parseGPUIndexFromOID(oid string) int {
    // Last component of OID is the GPU index
    parts := strings.Split(oid, ".")
    idx, _ := strconv.Atoi(parts[len(parts)-1])
    return idx
}

Files Modified

metrics_oids.go, for the new metric types + NVIDIA vendor OID mappings
device_profiles.go, for the 3 new profiles + GPUProfile struct + profile map entries
metrics_cycler.go, for the GPUMetrics struct, GPU data point generation, getter methods
snmp_handlers.go, for the GPU metric cases in getMetricValue()

Phase 2: SNMP Resource Files

2.1 Directory Structure

resources/
├── nvidia_dgx_a100/
│   ├── nvidia_dgx_a100_snmp_system.json    # sysDescr, sysObjectID, sysUpTime, sysContact, interfaces
│   ├── nvidia_dgx_a100_snmp_host.json      # Host Resources MIB (hrStorage, hrProcessor, hrDevice)
│   ├── nvidia_dgx_a100_snmp_gpu.json       # per-GPU objects under NVIDIA's PEN (nl6's own layout, static baselines)
│   ├── nvidia_dgx_a100_snmp_hcif.json      # IF-MIB ifXTable HC octet counter seed
│   ├── nvidia_dgx_a100_snmp_ifxtable.json  # IF-MIB ifXTable
│   ├── nvidia_dgx_a100_ssh.json            # SSH command/response pairs
│   └── nvidia_dgx_a100_api.json            # REST API endpoint responses
├── nvidia_dgx_h100/
│   ├── nvidia_dgx_h100_snmp_system.json
│   ├── nvidia_dgx_h100_snmp_host.json
│   ├── nvidia_dgx_h100_snmp_gpu.json
│   ├── nvidia_dgx_h100_snmp_hcif.json
│   ├── nvidia_dgx_h100_snmp_ifxtable.json
│   ├── nvidia_dgx_h100_ssh.json
│   └── nvidia_dgx_h100_api.json
└── nvidia_hgx_h200/
    ├── nvidia_hgx_h200_snmp_system.json
    ├── nvidia_hgx_h200_snmp_host.json
    ├── nvidia_hgx_h200_snmp_gpu.json
    ├── nvidia_hgx_h200_snmp_hcif.json
    ├── nvidia_hgx_h200_snmp_ifxtable.json
    ├── nvidia_hgx_h200_ssh.json
    └── nvidia_hgx_h200_api.json

2.2 SNMP System MIB (per flavor)

Standard MIB-II system group OIDs with GPU-server-appropriate values:

OID	Description	Example (DGX H100)
`1.3.6.1.2.1.1.1.0`	sysDescr	`NVIDIA DGX H100 - DCGM 3.3.0, Driver 535.129.03, CUDA 12.2`
`1.3.6.1.2.1.1.2.0`	sysObjectID	`1.3.6.1.4.1.5703.1.2.2`
`1.3.6.1.2.1.1.3.0`	sysUpTime	`234567800`
`1.3.6.1.2.1.1.4.0`	sysContact	`GPU Infrastructure Team`
`1.3.6.1.2.1.1.7.0`	sysServices	`72`

Plus IF-MIB interfaces (management NIC, InfiniBand/RoCE ports), Host Resources MIB entries (CPU, memory, storage).

2.3 SNMP GPU objects (nl6's own layout under NVIDIA's PEN)

Static OIDs under 1.3.6.1.4.1.5703 for per-GPU identity info. These object definitions are nl6's, not NVIDIA's: see the note at the top of this page.

OID Pattern	Description	Example
`.1.1.1.1.1.{gpu}`	GPU name	`NVIDIA H100 80GB HBM3`
`.1.1.1.1.2.{gpu}`	GPU UUID	`GPU-a1b2c3d4-...`
`.1.1.1.1.3.{gpu}`	GPU serial	`1234567890ABC`
`.1.1.1.1.4.{gpu}`	GPU PCI bus ID	`0000:07:00.0`
`.1.1.1.1.13.{gpu}`	Driver version	`535.129.03`
`.1.1.1.1.14.{gpu}`	CUDA version	`12.2`
`.1.1.1.1.15.{gpu}`	ECC errors (corrected)	`0`
`.1.1.1.1.16.{gpu}`	ECC errors (uncorrected)	`0`
`.1.1.1.1.17.{gpu}`	NVLink active links	`12`

Dynamic OIDs (cycling via MetricsCycler) are listed in Phase 1.1 above.

Files Created

resources/nvidia_dgx_a100/nvidia_dgx_a100_snmp_system.json
resources/nvidia_dgx_a100/nvidia_dgx_a100_snmp_host.json
resources/nvidia_dgx_a100/nvidia_dgx_a100_snmp_gpu.json
resources/nvidia_dgx_h100/nvidia_dgx_h100_snmp_system.json
resources/nvidia_dgx_h100/nvidia_dgx_h100_snmp_host.json
resources/nvidia_dgx_h100/nvidia_dgx_h100_snmp_gpu.json
resources/nvidia_hgx_h200/nvidia_hgx_h200_snmp_system.json
resources/nvidia_hgx_h200/nvidia_hgx_h200_snmp_host.json
resources/nvidia_hgx_h200/nvidia_hgx_h200_snmp_gpu.json

Phase 3: SSH Resources (`nvidia-smi` and `dcgmi` CLI)

3.1 Supported Commands

Each flavor's SSH resource file provides responses for these commands:

Command	Description
`nvidia-smi`	Default GPU summary table (all 8 GPUs, utilization, memory, temp, power)
`nvidia-smi -q -d MEMORY`	Memory details (used/free/total per GPU)
`nvidia-smi -q -d UTILIZATION`	Utilization details (GPU%, memory controller%, encoder%, decoder%)
`nvidia-smi -q -d TEMPERATURE`	Temperature details (current, shutdown, slowdown thresholds)
`nvidia-smi -q -d POWER`	Power details (draw, limit, default limit, min/max limits)
`nvidia-smi topo -m`	GPU topology matrix (NVLink connections)
`dcgmi discovery -l`	DCGM discovered GPUs
`dcgmi health -c`	DCGM health check
`show version`	Linux OS version with NVIDIA driver info
`uname -a`	Linux kernel info
`hostname`	Hostname
`free -h`	System memory summary
`lscpu`	CPU info (AMD EPYC / Intel Xeon for DGX)
`lspci \| grep -i nvidia`	PCI device listing for GPUs

Fourteen commands per flavor. Nothing else is answered.

3.2 Example `nvidia-smi` Output (H100)

Fri Mar 14 12:00:00 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 535.129.03   Driver Version: 535.129.03   CUDA Version: 12.2                 |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                  Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap  |           Memory-Usage | GPU-Util  Compute M. |
|=========================================+========================+======================|
|   0  NVIDIA H100 80GB HBM3          On  | 00000000:07:00.0   Off |                    0 |
| N/A   42C    P0             320W / 700W  |   45312MiB / 81920MiB  |     67%      Default |
|-----------------------------------------+------------------------+----------------------|
|   1  NVIDIA H100 80GB HBM3          On  | 00000000:0B:00.0   Off |                    0 |
| N/A   39C    P0             285W / 700W  |   38400MiB / 81920MiB  |     54%      Default |
...  (GPUs 2-7)
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage       |
|=========================================================================================|
|    0   N/A  N/A     12345      C   python3                                     44800MiB |
|    1   N/A  N/A     12346      C   python3                                     37888MiB |
...
+-----------------------------------------------------------------------------------------+

Files Created

resources/nvidia_dgx_a100/nvidia_dgx_a100_ssh.json
resources/nvidia_dgx_h100/nvidia_dgx_h100_ssh.json
resources/nvidia_hgx_h200/nvidia_hgx_h200_ssh.json

Phase 4: REST API Resources (DCGM HTTP Endpoints)

4.1 API Endpoints

Each device exposes a REST API on its APIPort with these endpoints:

Method	Path	Description
`GET`	`/api/v1/gpu/status`	Overall GPU cluster health and summary
`GET`	`/api/v1/gpu/devices`	List of all GPUs with IDs, names, UUIDs
`GET`	`/api/v1/gpu/topology`	NVLink/NVSwitch topology
`GET`	`/api/v1/dcgm/health`	DCGM health check results
`GET`	`/api/v1/dcgm/config`	DCGM configuration
`GET`	`/api/v1/system/info`	System info (OS, driver, CUDA, hostname)
`GET`	`/api/v1/system/memory`	System memory stats

Seven endpoints per flavor. There is no per-GPU /devices/{id} path.

4.2 Example Response: `/api/v1/gpu/status`

{
  "status": "healthy",
  "gpu_count": 8,
  "driver_version": "535.129.03",
  "cuda_version": "12.2",
  "dcgm_version": "3.3.0",
  "gpus": [
    {
      "index": 0,
      "name": "NVIDIA H100 80GB HBM3",
      "uuid": "GPU-a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "pci_bus_id": "00000000:07:00.0",
      "temperature_gpu": 42,
      "utilization_gpu": 67,
      "utilization_memory": 55,
      "memory_used_mib": 45312,
      "memory_total_mib": 81920,
      "power_draw_w": 320,
      "power_limit_w": 700,
      "fan_speed_pct": 0,
      "sm_clock_mhz": 1980,
      "mem_clock_mhz": 2619,
      "ecc_errors_corrected": 0,
      "ecc_errors_uncorrected": 0,
      "health": "healthy"
    }
  ]
}

4.3 Example Response: `/api/v1/gpu/topology`

{
  "gpu_count": 8,
  "nvlink_version": 4,
  "nvswitch_count": 4,
  "topology_matrix": [
    { "gpu": 0, "connections": [
      { "peer_gpu": 1, "type": "NVLink", "bandwidth_gbps": 900 },
      { "peer_gpu": 2, "type": "NVLink", "bandwidth_gbps": 900 },
      ...
    ]},
    ...
  ]
}

Files Created

resources/nvidia_dgx_a100/nvidia_dgx_a100_api.json
resources/nvidia_dgx_h100/nvidia_dgx_h100_api.json
resources/nvidia_hgx_h200/nvidia_hgx_h200_api.json

Phase 5: Integration with Simulator Framework

5.1 Resource Loading (`resources.go`)

No changes needed. The existing directory-based loading and merging already handles new resource directories automatically. The loadSpecificResourcesFromDir() function will pick up all JSON files in each NVIDIA directory and merge SNMP + SSH + API resources.

5.2 Device Type Detection (`resources.go`)

Add NVIDIA detection to getDeviceTypeFromName():

} else if strings.Contains(nameLower, "nvidia") || strings.Contains(nameLower, "dgx") || strings.Contains(nameLower, "hgx") {
    return "NVIDIA GPU Server"
}

5.3 Round-Robin Registration (`types.go`)

Add the 3 NVIDIA resource files to RoundRobinDeviceTypes:

var RoundRobinDeviceTypes = []string{
    // ... other types ...
    "nvidia_dgx_a100.json",
    "nvidia_dgx_h100.json",
    "nvidia_hgx_h200.json",
}

The shipped list has 29 device types in round-robin mode.

5.4 API Resource Merging (`resources.go`)

Update loadResourcesFromDir() and loadSpecificResourcesFromDir() to also merge API resources (currently only merges SNMP and SSH):

resources.API = append(resources.API, partResources.API...)

Initialize the API slice:

resources := &DeviceResources{
    SNMP: make([]SNMPResource, 0),
    SSH:  make([]SSHResource, 0),
    API:  make([]APIResource, 0),  // Add this
}

Files Modified

resources.go, for the NVIDIA device type detection + API merging in directory loader
types.go, for the 3 new entries in RoundRobinDeviceTypes

Phase 6: Build and Verify

Verify compilation with cd go && go build ./nl6/.
Verify there are no issues with cd go && go vet ./nl6/.
Start the simulator and create a single NVIDIA DGX H100 device
Verify SNMP walk returns GPU OIDs with cycling values
Verify SSH nvidia-smi returns formatted output
Verify REST API /api/v1/gpu/status returns GPU data
Create devices in round-robin mode. Verify NVIDIA types appear among the 29 types
Verify web UI shows "NVIDIA GPU Server" as device type

File Change Summary

New Files (21 resource JSON files)

resources/nvidia_dgx_a100/
    nvidia_dgx_a100_snmp_system.json
    nvidia_dgx_a100_snmp_host.json
    nvidia_dgx_a100_snmp_gpu.json
    nvidia_dgx_a100_snmp_hcif.json
    nvidia_dgx_a100_snmp_ifxtable.json
    nvidia_dgx_a100_ssh.json
    nvidia_dgx_a100_api.json

resources/nvidia_dgx_h100/
    nvidia_dgx_h100_snmp_system.json
    nvidia_dgx_h100_snmp_host.json
    nvidia_dgx_h100_snmp_gpu.json
    nvidia_dgx_h100_snmp_hcif.json
    nvidia_dgx_h100_snmp_ifxtable.json
    nvidia_dgx_h100_ssh.json
    nvidia_dgx_h100_api.json

resources/nvidia_hgx_h200/
    nvidia_hgx_h200_snmp_system.json
    nvidia_hgx_h200_snmp_host.json
    nvidia_hgx_h200_snmp_gpu.json
    nvidia_hgx_h200_snmp_hcif.json
    nvidia_hgx_h200_snmp_ifxtable.json
    nvidia_hgx_h200_ssh.json
    nvidia_hgx_h200_api.json

Modified Files (6 Go files)

File	Changes
`metrics_oids.go`	8 MetricOIDType constants + 3 NVIDIA vendor OID map entries (64 GPU OIDs each)
`device_profiles.go`	GPUProfile struct + 3 GPU device profiles + profile map entries
`metrics_cycler.go`	GPUMetrics struct + GPU data generation in NewMetricsCycler + 9 GPU getter methods
`snmp_handlers.go`	GPU metric cases in getMetricValue() + parseGPUIndexFromOID helper
`resources.go`	NVIDIA device type detection + API merging in directory loaders
`types.go`	3 new entries in RoundRobinDeviceTypes

No Changes Required

File	Why
`device.go`	MetricsCycler already initialized per device; GPU init keyed on resource file
`manager.go`	Resource caching and device lifecycle unchanged
`snmp_server.go`	Protocol handling unchanged
`snmp_encoding.go`	Encoding unchanged
`ssh.go`	Command matching unchanged
`api.go`	API server already handles arbitrary endpoints from resource files
`web.go`	Web UI device listing works with new device types automatically
`web_routes.go`	Routes unchanged
`web/index.html`	Device type dropdown auto-populated from resources
`web/app_ui.js`	Renders device_type from API response automatically
`web/app_api.js`	API client unchanged

Overview​

What DCGM Exposes in the Real World​

Scope​

Device Flavors (3 resource profiles)​

GPU-Specific Metric Types (new)​

Phase 1: GPU Metric Types and Cycler Extension​

1.1 New Metric OID Types (metrics_oids.go)​

1.2 GPU Device Profile (device_profiles.go)​

1.3 GPU Metrics in MetricsCycler (metrics_cycler.go)​

1.4 Getter Methods for GPU Metrics​

1.5 SNMP Handler Extension (snmp_handlers.go)​

Files Modified​

Phase 2: SNMP Resource Files​

2.1 Directory Structure​

2.2 SNMP System MIB (per flavor)​

2.3 SNMP GPU objects (nl6's own layout under NVIDIA's PEN)​

Files Created​

Phase 3: SSH Resources (nvidia-smi and dcgmi CLI)​

3.1 Supported Commands​

3.2 Example nvidia-smi Output (H100)​

Files Created​

Phase 4: REST API Resources (DCGM HTTP Endpoints)​

4.1 API Endpoints​

4.2 Example Response: /api/v1/gpu/status​

4.3 Example Response: /api/v1/gpu/topology​

Files Created​

Phase 5: Integration with Simulator Framework​

5.1 Resource Loading (resources.go)​

5.2 Device Type Detection (resources.go)​

5.3 Round-Robin Registration (types.go)​

5.4 API Resource Merging (resources.go)​

Files Modified​

Phase 6: Build and Verify​

File Change Summary​

New Files (21 resource JSON files)​

Modified Files (6 Go files)​

No Changes Required​

Overview

What DCGM Exposes in the Real World

Scope

Device Flavors (3 resource profiles)

GPU-Specific Metric Types (new)

Phase 1: GPU Metric Types and Cycler Extension

1.1 New Metric OID Types (`metrics_oids.go`)

1.2 GPU Device Profile (`device_profiles.go`)

1.3 GPU Metrics in MetricsCycler (`metrics_cycler.go`)

1.4 Getter Methods for GPU Metrics

1.5 SNMP Handler Extension (`snmp_handlers.go`)

Files Modified

Phase 2: SNMP Resource Files

2.1 Directory Structure

2.2 SNMP System MIB (per flavor)

2.3 SNMP GPU objects (nl6's own layout under NVIDIA's PEN)

Files Created

Phase 3: SSH Resources (`nvidia-smi` and `dcgmi` CLI)

3.1 Supported Commands

3.2 Example `nvidia-smi` Output (H100)

Files Created

Phase 4: REST API Resources (DCGM HTTP Endpoints)

4.1 API Endpoints

4.2 Example Response: `/api/v1/gpu/status`

4.3 Example Response: `/api/v1/gpu/topology`

Files Created

Phase 5: Integration with Simulator Framework

5.1 Resource Loading (`resources.go`)

5.2 Device Type Detection (`resources.go`)

5.3 Round-Robin Registration (`types.go`)

5.4 API Resource Merging (`resources.go`)

Files Modified

Phase 6: Build and Verify

File Change Summary

New Files (21 resource JSON files)

Modified Files (6 Go files)

No Changes Required