Storage for telecom OSS/BSS servers: arrays, drive tiers and spares
Telecom OSS/BSS servers need three storage tiers: NVMe SSD for subscriber, billing and CRM databases; SAS SSD for mediation, alarms and indexes; and 7,200 rpm nearline HDD or LTO tape for CDR and performance archives. A 20-million-subscriber model lands at 74 TB usable, or 101 TB raw after RAID 6 parity and a 20% reserve.
What storage workloads does an OSS/BSS stack create?
ITU-T M.3010 defines the telecommunications management network, and the FCAPS split that most OSS products follow comes from the X.700 framework: fault, configuration, accounting, performance and security.
Charging generates the largest stream. 3GPP TS 32.240 defines the charging architecture, and 3GPP TS 32.298 lists the CDR parameters that mediation or billing later parses. Records are written once, read during rating, and read again during dispute or audit work.
Performance management produces the second stream. 3GPP TS 32.401 defines the PM concept; a 15-minute granularity period yields 96 samples per counter per day. Ten thousand network elements with 2,000 counters each generate 20 million samples every 15 minutes, or 1.92 billion per day. At 8 bytes per sample that is 15.4 GB of raw counter data per day.
Fault and inventory data sit beside the other streams: small random writes, indexed reads, and a store that behaves like a database. ITU-T X.733 defines five perceived severities: critical, major, minor, warning and indeterminate.
Billing and CRM hold the state that must survive everything else. A 4 KB subscriber record across 20 million subscribers is 80 GB, and production copies, indexes and redo logs multiply it by three. Month-end rating turns into sequential CDR reads and random probes into rating tables.
The mix explains why one array type rarely covers an estate: database traffic rewards latency, CDR and PM traffic rewards throughput, alarm traffic rewards small random writes.
How much capacity and IOPS does a 20-million-subscriber OSS/BSS need?
Capacity planning starts from three multipliers: records per subscriber per day, bytes per record, and days of retention. The table works a 20-million-subscriber example at 13 months, a period that covers one annual audit cycle and one month of overlap, and adds the peak IOPS and latency each workload places on the array.
| Workload | Unit size | Daily volume | 13-month raw | Peak IOPS (design target) | Latency target | Planning note |
|---|---|---|---|---|---|---|
| CDR stream | 800 B per record | 200 million records/day = 160 GB/day | 63.2 TB | 70 at 64 KB | 2 ms | 10 records per subscriber per day; 3GPP TS 32.298 parameter set |
| PM counters | 8 B per sample | 1.92 billion samples/day = 15.4 GB/day | 6.1 TB | 270 at 64 KB | 5 ms | 10,000 NEs x 2,000 counters x 96 samples; 3GPP TS 32.401 |
| Alarm and event store | 2 KB per alarm | 200,000 alarms/day = 400 MB/day | 0.16 TB | 231 at 2 KB | 1 ms | Five X.733 severities, indexed; 100x storm |
| Subscriber and CRM database | 4 KB per subscriber | 80 GB per copy | 0.24 TB | 100,000 at 8 KB | 200 microseconds | Three copies including redo and standby |
| Billing extracts and reports | 80 GB per run | 160 GB/month | 2.1 TB | 50,000 at 8 KB | 2 ms | Month-end and mid-cycle runs |
| Logs and traces | 5 GB/day | 5 GB/day | 2.0 TB | 50 at 16 KB | 10 ms | Application and audit logs |
| Total | 73.8 TB | Before RAID and free-space reserve |
RAID 6 on a 24-bay shelf, where 22 of 24 drives carry data, plus a 20% reserve takes 73.8 TB to 101 TB raw (73.8 / 0.917 / 0.8 = 100.6).
Two inputs dominate the total. Retention is the first: extending CDR storage from 13 to 25 months adds 58 TB, because the CDR row dominates the sum. The reduction ratio is the second. The OceanStor Dorado 3000 V6 datasheet quotes 3:1 data reduction with inline dedupe and compression; CDR and PM records carry unique timestamps and values, and CDRs also carry subscriber identifiers, so duplicate blocks are rare. A plan that applies 3:1 to those stores lands short of the real requirement.
Smaller operators scale linearly. Two million subscribers with a tenth of the network elements and the same record sizes need 7.4 TB usable, or 10.1 TB raw under the same RAID and reserve assumptions.
Which drive tier fits each OSS/BSS workload?
Six drive classes cover the estate. The interface ceiling is fixed by the standard; the drive figure comes from a named datasheet.
| Tier | Interface ceiling | Documented drive figure | OSS/BSS role |
|---|---|---|---|
| NVMe TLC, U.2 or E3.S | PCIe 5.0 x4, 15.75 GB/s | Samsung PM1743: up to 14,000 MB/s sequential read, 15.36 TB | Subscriber, billing and CRM databases |
| SAS SSD | SAS 12 Gb/s, 1,200 MB/s per lane | Dual-port 2.5-inch | Mediation, indexes, alarm store |
| SATA SSD | 6 Gb/s, 600 MB/s | 2.5-inch | Boot, small OSS applications |
| 15K SAS HDD | 12 Gb/s, 1,200 MB/s per lane | 2.5-inch, 15,000 rpm | Legacy alarm and inventory databases |
| 7.2K nearline HDD | 12 Gb/s, 1,200 MB/s per lane | Ultrastar DC HC550 14 TB: 262 MB/s sustained, 512 MB cache | CDR and PM archive |
| LTO-9 tape | SAS 12 Gb/s or FC | 400 MB/s native, 18 TB per cartridge | Regulatory and long-term archive |
Interface ceilings are arithmetic. PCIe 5.0 moves 32 GT/s per lane with 128b/130b encoding, or 3.94 GB/s per lane and 15.75 GB/s across four lanes; PCIe 4.0 halves that to 7.88 GB/s. SAS 12 Gb/s delivers 1,200 MB/s per lane after 8b/10b encoding, and SAS 24 Gb/s uses 128b/150b. SATA 6 Gb/s stops at 600 MB/s.
U.2, standardised as SFF-8639, carries four PCIe lanes or two SAS lanes in a 2.5-inch bay, so one 24-bay chassis takes NVMe or SAS drives. The NVMe base specification allows 65,535 I/O submission queues with 65,536 entries each; a SAS drive carries one queue with a 256-command tag space per port. The difference matters for subscriber lookups, much less for a CDR file read sequentially.
A 7,200 rpm spindle completes a revolution in 8.33 ms, so average rotational latency is 4.17 ms; a 15,000 rpm spindle averages 2.0 ms. With seek time, the two land near 80 and 200 random IOPS.
QLC flash belongs in read-mostly archives. A CDR store writes its full capacity every retention period, and QLC program/erase cycles are fewer than TLC; compare drive writes per day under JEDEC JESD219A.
How many IOPS and what latency should the array deliver?
Array datasheets quote peak numbers with cache hits. The Infortrend EonStor DS 4000U datasheet lists up to 1,000K IOPS and 11 GB/s for the 2U 24-bay DS 4024U, with 8 GB to 128 GB of DDR4 cache per controller. A billing batch reading 8 KB blocks at 50,000 IOPS consumes 400 MB/s and leaves most of the 11 GB/s unused.
Little’s law converts latency into concurrency: outstanding I/O equals IOPS multiplied by latency in seconds. At 100,000 IOPS and 200 microseconds, the array must hold 20 commands in flight. At 50,000 IOPS and 2 milliseconds, it must hold 100.
Queue depth decides whether that second figure is reachable. An NVMe drive with 65,535 queues has more headroom than a SAS drive with one 256-command queue per port, even at the same array-level IOPS.
Workload peaks are easier to bound than averages. The 160 GB/day CDR stream averages 1.85 MB/s; if 10% of the day’s records arrive in the busiest hour, that hour carries 4.4 MB/s. Replaying a day of PM files inside one 15-minute window needs 17.1 MB/s. Neither stresses a 25GbE link; both stress the metadata path if millions of small files are created per cycle.
Alarms are the opposite case. 200,000 alarms per day average 2.3 per second, and a storm at 100 times the average reaches 231 per second. Each alarm is a small indexed write followed by correlation reads, so the array sees 2 KB to 8 KB random I/O, the profile that benefits most from NVMe and protected write-back cache.
Latency targets follow the same split. Subscriber lookups and rating calls target 100 to 200 microseconds; billing batch and mediation tolerate 1 to 2 milliseconds; CDR and PM archive scans tolerate 10 milliseconds. All-flash meets the first two; nearline HDD meets the third and stalls on the first.
How should bays, RAID and expansion be laid out?
Bay count sets the failure domain. A 24-bay RAID 6 set survives two drive failures and loses 8.3% of raw capacity to parity; the same shelf with one hot spare loses 12.5% and keeps a rebuild target ready. A 12-bay 3.5-inch shelf loses 16.7% to RAID 6 parity, and adding a spare takes the loss to 25%. Wide RAID sets raise rebuild time and the chance of a second failure during rebuild, so 24 bays is the practical ceiling for 7.68 TB SSDs and 12 bays for 20 TB HDDs.
| Shelf | Bays | Drive | RAID | Data drives | Raw capacity | Usable capacity |
|---|---|---|---|---|---|---|
| 2U NVMe | 24 x 2.5-inch U.2 | 7.68 TB | RAID 6 | 22 | 184 TB | 169 TB |
| 2U NVMe with hot spare | 24 x 2.5-inch U.2 | 7.68 TB | RAID 6 + 1 spare | 21 | 184 TB | 161 TB |
| 2U nearline | 12 x 3.5-inch | 20 TB | RAID 6 | 10 | 240 TB | 200 TB |
| 2U nearline with hot spare | 12 x 3.5-inch | 20 TB | RAID 6 + 1 spare | 9 | 240 TB | 180 TB |
| 2U NVMe, RAID 10 | 24 x 2.5-inch U.2 | 7.68 TB | RAID 10 | 12 mirrored pairs | 184 TB | 92 TB |
Rebuild arithmetic is unforgiving at these capacities. A 7.68 TB SSD rebuilt at 500 MB/s takes 4.3 hours. A 20 TB HDD rebuilt at 262 MB/s takes 21.2 hours; at half that rate, 42.4 hours. That number belongs in the risk register beside any 12-bay nearline shelf.
RAID-TP adds a third parity drive. The Dorado 3000 V6 datasheet states that RAID-TP tolerates three simultaneous SSD failures, which changes the calculus for large SSDs. RAID 10 stays the right choice for redo logs: half the raw capacity, but rebuilds copy a mirror instead of recomputing parity.
Expansion paths differ by product. The EonStor DS 4000U scales to 448 drives through JB 3012, JB 3016, JB 3024B, JB 3060L and JB 3090 enclosures, 18.7 shelves of 24 drives. The CS 4000U scales to 144 nodes and 100 PB, with 120 drives per node through JBOD. The Dorado 3000 V6 tops out at 1,200 SSDs and 16 controllers, the 5000 V6 at 3,200, the 6000 V6 at 4,800.
Which interfaces carry OSS/BSS storage traffic?
Fibre Channel remains the default block transport in carrier OSS/BSS estates. A 32G FC port carries 3,200 MB/s, a 16G port 1,600 MB/s, and dual fabrics give two independent paths. The DS 4000U offers 16G FC x4 or 32G FC x2/x4 host boards; the Dorado 3000 V6 lists FC, iSCSI, NFS, CIFS, FC-NVMe and NVMe over RoCE.
Ethernet options price lower per port. 25GbE carries 3,125 MB/s, 10GbE 1,250 MB/s, and the DS 4000U offers 10GbE SFP+ x2 or 25GbE SFP28 x2. NVMe over RoCE is deployed with priority flow control under IEEE 802.1Qbb and ECN marking to keep the pause domain small; a network without PFC risks dropping RoCE frames under congestion.
Port count follows from array throughput. 11 GB/s needs at least four 32G FC ports (12.8 GB/s) before protocol overhead, and eight if the design keeps two fabrics at full rate. 1,000K IOPS at 8 KB blocks is 8 GB/s of demand, so the four-port minimum applies. A two-port HBA caps a server at 6.4 GB/s.
Multipath software belongs in the bill of materials from the start. A single HBA, switch and controller port turn a firmware update into an outage. The 25G and 100G optical modules guide maps reach and fibre type, and the 25G SFP28 transceiver page lists the catalogue parts.
What spare-parts stock keeps an OSS/BSS array running?
Drive failures follow the bathtub curve: high early, flat middle, rising wear-out. A 72-hour burn-in screens early failures and applies to refurbished stock. SMART attribute 5 (reallocated sectors), 187 (reported uncorrectable), 197 (current pending sectors) and 198 (offline uncorrectable) give the warning; SAS drives use the grown defect list.
One cold spare per 24-bay shelf is 4.2% of the population; two are 8.3%. One per 12-bay shelf is 8.3%, two are 16.7%. A 20 TB nearline shelf warrants two spares because a 21.2-hour rebuild leaves a long window; a 7.68 TB NVMe shelf can run with one.
A stock list adds a controller per eight arrays, one power supply per four shelves, one fan tray per shelf, one cache-protection module per controller. The DS 4000U backs up cache with a supercapacitor plus flash module, and supercapacitors age; an aged module is a scheduled replacement.
Firmware and sector format matching matter as much as the part number. A drive with different firmware may refuse to join a pool or join at reduced performance, and a 512e drive in a 4Kn pool can be rejected by the array or forced into emulation. Record the approved model, firmware revision and sector format; buy from the same qualification batch where possible.
Cold spares need a home. A drive stored for years without power can develop stiction and bearing problems; rotate cold spares through the array or power them quarterly. The refurbished telecom equipment inspection checklist covers receiving checks, and the export compliance checklist covers documentation for encrypted or data-bearing drives.
Which catalogue arrays fit a pilot, a regional site and a national deployment?
| Model | Type | Maximum drives | Cache | LUNs | Host interfaces | Deployment fit |
|---|---|---|---|---|---|---|
| OceanStor Dorado 3000 V6 | All-flash SAN | 1,200 SSDs | 128 GB to 1,536 GB | 8,192 | FC, iSCSI, NFS, CIFS, FC-NVMe, NVMe over RoCE | Regional OSS/BSS |
| OceanStor Dorado 5000 V6 | All-flash SAN | 3,200 SSDs | 256 GB to 8 TB | 16,384 | Above plus NDMP | National billing and CRM |
| OceanStor Dorado 6000 V6 | All-flash SAN | 4,800 SSDs | 1 TB to 24 TB | 32,768 | Above plus NDMP | Large billing database |
| EonStor DS 4000U | NVMe SAN | 24 per shelf, 448 maximum | 8 GB to 128 GB per controller | Not stated | 16/32G FC, 10/25GbE, 12G SAS | CDR and PM block storage |
| EonStor CS 4000U | Scale-out NAS | 14 per node, 120 with JBOD, 144 nodes | 64 GB to 256 GB per node | Not stated | 10/25GbE | PM and log archive |
| FusionServer 2288H V7 | 2U rack server | 45 x 2.5-inch or 28 NVMe U.2 | System memory only | Not applicable | PCIe 5.0, OCP 3.0 | Pilot and hyperconverged OSS |
| FusionServer 2488H V7 | 2U 4-socket server | 8/25 front 2.5-inch or 4 + 8 NVMe | System memory only | Not applicable | PCIe 5.0, OCP 3.0 | Billing database host |
The 101 TB raw target fits a single 24-bay NVMe shelf at 7.68 TB per drive: 14 drives give 107.5 TB raw, a 14-drive RAID 6 set gives 92.2 TB usable, and the 20% reserve brings that to 73.7 TB. Ten of the 24 bays stay empty for growth.
Scale-out NAS fits the file-based side of OSS. PM files, alarm logs and trace archives arrive as files, and the CS 4000U puts 14 U.2 SSDs in each 2U node with a 3-node cluster rated at 18 GB/s. Its 144-node, 100 PB ceiling makes it a consolidation target for regional sites.
Server-local storage has a narrower role. The 2288H V7 holds 45 x 2.5-inch SAS/SATA/NVMe drives or 28 NVMe U.2 drives in 2U, enough for a pilot or a hyperconverged cluster. The 2488H V7 trades drive count for four CPU sockets and 64 DDR5 slots, a shape that suits an in-memory billing database with a local mirror.
The OceanStor Dorado 3000 V6, 5000 V6, 6000 V6, EonStor DS 4000U, EonStor CS 4000U, FusionServer 2288H V7 and 2488H V7 pages carry full specification sheets, and the equipment catalogue lists tested units.
Frequently asked questions
Can a single 2U server run OSS/BSS storage?
For a pilot, yes. A FusionServer 2288H V7 with 28 NVMe U.2 drives holds a complete OSS stack in 2U. Production adds two constraints: no dual-controller failover, and a chassis failure takes every drive with it. A dual-controller shelf keeps drives reachable when one controller dies.
Is RAID 6 safe with 20 TB drives?
RAID 6 survives two failures, and the rebuild is the weak point. A 20 TB drive rebuilt at 262 MB/s takes 21.2 hours; at half that rate, 42.4 hours. During that window a second failure costs the set. Three mitigations: keep the RAID set at 12 drives or fewer, hold two cold spares, and keep a second copy of the CDR archive on tape. RAID-TP, available on the Dorado 3000 V6, adds a third parity drive and tolerates three simultaneous SSD failures.
Do CDRs deduplicate or compress well?
Poorly. A CDR carries a timestamp, subscriber identity, cell identity, service type and counters, and the combination is close to unique per record. Dedupe finds few repeated blocks, and general-purpose compression yields modest savings on numeric fields. Plan CDR capacity at 1:1 and treat the datasheet 3:1 reduction ratio as a mixed-workload figure that applies to other data.
How long should CDRs stay online?
3GPP TS 32.240 sets no retention period; the regulator and the dispute process do. Thirteen months covers an annual audit plus a month of overlap, and 25 months doubles the CDR footprint from 63.2 TB to 121.6 TB. A common compromise keeps 90 days on SSD for rating, the rest of the 13 months on nearline HDD, and older records on LTO-9 tape.
Does NVMe over RoCE need a lossless network?
Yes. RoCE is normally deployed with priority flow control (IEEE 802.1Qbb) to stop drops and ECN marking to keep pause frames local to the congested link. A lossy Ethernet network risks retransmission and stalls under load. Where the network cannot be configured for PFC, iSCSI over 25GbE is the simpler block path, and the DS 4000U offers both.
Can the EonStor DS 4000U take third-party U.2 drives?
The datasheet states that 2.5-inch U.2 NVMe SSDs must be purchased from Infortrend; a third-party drive may fit the bay and still be rejected by the controller or excluded from the warranty.
How many 32G FC ports does an 11 GB/s array need?
Four. A 32G FC port carries 3,200 MB/s, so four ports carry 12.8 GB/s before protocol overhead, just above the 11 GB/s array rating. Eight ports keep two fabrics at full rate and survive one fabric failure. A two-port HBA caps the host at 6.4 GB/s.
Need a specific part?
Send the part number and quantity — we confirm condition, availability and price before you commit.