Skip to main content

Server

The Server page shows deep detail for the currently picked node — system specs, GPU/DPU/NIC composition, BIOS, kernel, and PCIe topology. Pick the active node from the sidebar's node switcher.

Server detailServer detail

Sections​

System Information​

Hostname, OS, kernel, architecture, manufacturer, product name, serial, uptime, and last boot time — useful when filing a support ticket.

Processor & Memory​

CPU model, core/thread count, live CPU and memory usage, and total installed memory.

BIOS Settings​

A panel of the BIOS-controlled platform features that matter for GPU/DPU workloads, inferred entirely from the agent's sysfs / procfs / MSR data — no BMC required:

  • Above 4G Decoding — required for large-BAR GPUs and DPUs
  • Resizable BAR — with the largest resizable VRAM BAR detected
  • SR-IOV — with PF count and total VF capacity
  • CPU Virtualization (VT-x / AMD-V) — the true BIOS state read from the feature-control MSR, not just the /proc/cpuinfo capability flag
  • IOMMU / VT-d / AMD-Vi — reports the firmware toggle (ACPI DMAR/IVRS table present) separately from kernel activation, so "enabled in BIOS but the kernel isn't using it" (no intel_iommu=on / amd_iommu=on) is distinguishable
  • PCIe ASPM Policy — active link power-management policy

Each reads Enabled / Disabled (or the ASPM policy name). Data comes from GET /api/v1/nodes/:id/pcie-bios.

PCIe Devices​

A combined view of every GPU and NIC the agent discovered. Four tabs share the section header:

  • Device List — flat table with type, name, BDF, NUMA node, and the PCIe path from root complex to the device.

  • Physical Slots — chassis's physical PCIe slot inventory derived from the agent's dmidecode + lspci data (no BMC required). Slots whose SMBIOS labels collide are disambiguated as (×N) or #N so each row maps unambiguously to a single endpoint.

  • PCIe Tree — hierarchical view of the host's PCI tree with every GPU and NIC labeled.

  • Hop Heatmap — GPU ↔ NIC hop-distance matrix, invaluable for tuning GPUDirect RDMA:

    • hops = 1 — GPU and NIC share a PCIe switch (best)
    • hops = 2 — both attach to the same root complex but cross the CPU
    • hops ≥ 3 — traffic crosses inter-CPU links (slowest)

    Pin RDMA streams to the closest NIC for each GPU using this map.

Storage Health​

A summary strip with disk count, used capacity, worst SSD life-used, and total media errors. Drill into Hardware → Storage for per-device SMART data.

Cluster-Wide View​

The Server page is per-node. For the whole-fleet list, see Cluster → Nodes.