Transceiver dalam rangkaian storan
Pasukan kami sedang menyiapkan terjemahan. Artikel ini buat sementara waktu belum tersedia dalam bahasa Melayu dan dipaparkan dalam bahasa Inggeris.
Storage is the environment where optics are least forgiving: arrays, HBAs and SAN switches all validate modules, links must be lossless, and a marginal transceiver shows up as application latency rather than a down port. This page covers how transceivers are used across storage networks and what to watch for.
Where the optics sit
| Location | Typical optics | Notes |
|---|---|---|
| Array front-end ports | 16/32GFC SFP+/SFP28; 25/100G Ethernet SFP28/QSFP28 | array vendors publish qualified-parts lists and often ship optics pre-installed |
| Host HBAs | 16/32GFC SFP+/SFP28 | HBAs are shipped with optics; third-party modules generally work but complicate support |
| SAN switch ports and ISLs | SW for in-room, LW/ELW for distance | the strictest validation — see below |
| IP storage NICs (iSCSI, NVMe/TCP, RoCE) | 25/100G Ethernet optics or DAC | standard Ethernet rules apply |
| Replication links | LW/ELW, or FC/Ethernet over DWDM | distance and latency budgets |
Vendor validation everywhere
Storage platforms check module identity on three levels:
- SAN switches — many show a module invalid port state for unqualified optics and keep the port offline (FC optics).
- Arrays — controllers may log an unsupported-SFP alert or refuse to bring a port up.
- HBAs — usually tolerant, but vendor support may require qualified optics before investigating a fault.
The practical policy operators converge on: qualified optics on array and switch ports, third-party where tolerated, and a coding tool to make a replacement present the identity a strict port expects (see Vendor lock).
Lossless means "slow", not "down"
FC and RoCE do not drop frames. A degraded optic therefore produces CRC errors, credit stalls (slow-drain) and rising latency on the LUNs behind it — a storage performance ticket, not a network outage. The diagnostic path is the same as for any link (levels → clean → swap → OTDR), but you must start from the application symptom and work back to the port.
Distance and replication
- Synchronous replication is bounded by latency — in practice up to roughly 100 km — which is why metro DR sites use LW/ELW optics or FC over DWDM.
- Long ISLs need extra buffer credits as well as optical budget; see SAN fabrics & distance.
Monitoring
Trend DDM on every ISL and array port: Rx power drift and Tx bias growth are the early warning of a laser that will fail on a Friday night. Baseline each optic at installation.
In CodingBox
Bench-read every optic before it goes into a storage network: confirm FC or Ethernet rate codes, wave type and vendor identity, log DDM as the baseline, and — for strict ports — code the module to the qualified identity in the EEPROM editor with the original preserved in the code database.
What lossless Ethernet demands of every link — PFC, ECN, FEC, zero CRC: Lossless Ethernet for storage; the transports that ride on it: NVMe over Fabrics.