SB9
SB9 is the wired-only sandbox at WINLAB (Route 1, North Brunswick, NJ), purpose-built for programmable-NIC, FPGA, GPU, and software-defined- networking experimentation. There are no radios on these nodes — the intent is high-speed deterministic networking and accelerated compute.
SB9 is the right pick whenever you need: high-speed (25 / 100 GbE) server-to-server interconnect, programmable data-plane experimentation on Xilinx Alveo / NetFPGA hardware, or GPU-backed back-end work without contending for the Grid’s SDR infrastructure.
Figure 1 — SB9 block diagram.
The 8 servers
Eight identical Dell PowerEdge R740 chassis: srv1.sb9.cosmos-lab.org
through srv8.sb9.cosmos-lab.org. Per-server baseline (authoritative,
from the omf-inventory scan of 2026-05-10):
| Component | Value |
|---|---|
| Mainboard | Dell PowerEdge R740 (BIOS 2.25.0) |
| CPU | 2× Intel Xeon Gold 6126 (12 cores / 24 threads per socket; 24 C / 48 T total) |
| RAM | 192 GB DDR4-2666 (24 DIMMs) |
| Disk | Samsung enterprise SATA SSD (~223 GB — mix of PM863a, SSD-860, SSD-883 across servers) |
| GPU | NVIDIA Tesla V100 PCIe 16 GB |
| FPGA — main accelerator | Xilinx Alveo U200 XDMA Platform (PCIe Gen3 ×16) |
| FPGA — SmartSSD | Xilinx SmartSSD bridge (storage accelerator) |
| Programmable NIC | 2× Xilinx Device 0100 Ethernet (NetFPGA-style 100 GbE on the U200 fabric) |
| Host NICs | 6× Mellanox ConnectX-5; eth2 is the active 25 GbE control link |
| BMC | Dell iDRAC9 (firmware 7.00.00.184) |
Per the orbit-lab.org SB9 reference, some nodes also carry additional accelerators / NICs:
node1-5,node1-6(=srv5,srv6) — Xilinx Alveo SN1022 SmartNIC (2× 100 GbE)node1-7,node1-8(=srv7,srv8) — NetFPGA 10G (4× 10 GbE)
Verify with lspci -nn | grep -iE 'xilinx|netfpga|mellanox' on the
specific node before designing topology — the slot population has
changed over time.
Switching fabric
SB9 has two top-of-rack switches for flexible data-plane experimentation:
| Switch | Ports | OS / Use |
|---|---|---|
| Dell Z9332F-ON | 32× 100 GbE (each breakable to 2×50, 2×40, 4×25, or 4×10) | Dell OS10. Management: in-band, out-of-band, or onboard BMC. |
| Edgecore Wedge100BF-32x | 32× 100 GbE (same breakout options) | Azure SONIC or OpenNetworkLinux. Management: in-band, out-of-band, or onboard BMC. |
Each server has its main 100 GbE links to both switches so you can:
- run experiments on one switch while the other carries the control plane
- A/B-test a SONIC build against Dell OS10 with the same physical wiring
- run a P4 or DASH data-plane on the Wedge while the Dell carries unmodified traffic
Data-switch connections
Full per-port mapping is documented on the orbit-lab.org SB9 reference. A sample of the Dell Z9332F-ON Port 1/1/N port assignment:
| Switch port | Connection | Speed |
|---|---|---|
| 1/1/1:1 | srv1 Mellanox ConnectX-5 Port 1 | 100 G |
| 1/1/2:1 | srv2 Mellanox ConnectX-5 Port 1 | 100 G |
| 1/1/3:1 | srv3 Mellanox ConnectX-5 Port 1 | 100 G |
| 1/1/4:1 | srv4 Mellanox ConnectX-5 Port 1 | 100 G |
| 1/1/5:1 | srv5 Mellanox ConnectX-5 Port 1 | 100 G |
| 1/1/6:1 | srv6 Mellanox ConnectX-5 Port 1 | 100 G |
| 1/1/7:1 | srv7 Mellanox ConnectX-5 Port 1 | 100 G |
| 1/1/8:1 | srv8 Mellanox ConnectX-5 Port 1 | 100 G |
| 1/1/9:1 | srv5 Alveo SN1022 Port 1 | 100 G |
| 1/1/10:1 | srv6 Alveo SN1022 Port 1 | 100 G |
| 1/1/11:1 – 1/1/18:1 | srv1…srv8 Alveo U200 Port 1 | 100 G |
| 1/1/20:1 … :4 | NetFPGA 10G ports on srv7/srv8 | 10 G |
| 1/1/21:1 … :4 | Mellanox ConnectX-4 Lx 25 G | 25 G |
Networking
SB9 was migrated from the legacy sb9.orbit-lab.org namespace to
sb9.cosmos-lab.org in the 2026 domain rename. Addressing after the cutover:
| Plane | Subnet | Notes |
|---|---|---|
| Control | 10.7.9.0/24 |
gateway 10.7.9.1; servers srv1–srv8 = 10.7.9.11–10.7.9.18 |
| Data1 | 10.8.9.0/24 |
high-speed experiment plane (25 / 100 GbE) |
| Data2 | 10.9.9.0/24 |
second experiment plane |
| Chassis mgmt (CM) | 10.1.109.0/24 |
unchanged — iDRAC/IPMI cm-srv1–cm-srv8 = 10.1.109.17–10.1.109.24 |
Console: internal 10.7.9.10 (consolec / console-internal), public DMZ
10.50.9.10 (console.sb9.cosmos-lab.org). Infrastructure switches on the
control plane: sw-da-01 10.7.9.100, sw-da-02 10.7.9.101,
sdn 10.7.9.102, sw-sb-09 10.7.9.253.
Chassis manager
iDRAC9 / IPMI over LAN on 10.1.109.17 – 10.1.109.24 (one per server).
Power control routed through omf-cmc. The omf tell and omf load
commands transparently use this chassis-management path.
Recent maintenance note (May 2026): all 8 iDRACs upgraded from 7.00.00.183 → 7.00.00.184 (Dell catalog’s current 14G release). Earlier versions had a chassis-power state-machine bug where the iDRAC would accept
chassis power onbut the chassis would not actually energize; recovery required a physical AC cycle. The 7.00.00.184 firmware prevents new occurrences; SSH access from the cmc user works end-to-end, andomf tell -a on / offhcycles behave normally.
Choosing SB9
| If your experiment needs… | Use SB9 |
|---|---|
| Programmable NIC / FPGA data plane (Alveo U200, NetFPGA, SN1022) | ✓ |
| GPU back-end (V100) for ML | ✓ |
| 25 / 100 GbE between hosts | ✓ |
| Dual-switch (Dell OS10 + SONIC) experiments | ✓ |
| Wireless / SDR | use SB1 / SB2 instead |
| Repeatable RF attenuation matrix | use SB4 instead |
Access
omf-tell -a on -t "srv1.sb9.cosmos-lab.org,srv2.sb9.cosmos-lab.org,…"
omf-load -i baseline.ndz -t <same>
omf-stat -t <same>
omf-tell -a offh -t <same>
SB9 does not have a system:topo:all alias by default — enumerate the
servers explicitly.