Glossary of Terms
0-9
- 3-2-1-1-0 rule
- Backup guideline: three copies on two media types, one offsite, one immutable or offline, and zero errors in restore tests.
- 3D NAND
- Flash memory built as vertical stacks of hundreds of cell layers, raising density without shrinking each cell further.
A
- Air gap
- Physical or logical separation of a data copy from any network, so attackers and ransomware cannot reach it.
B
- Bandwidth (throughput)
- How much data moves per second, measured in MB/s or GB/s; equals IOPS multiplied by the I/O size.
- Bit rot
- Silent corruption of stored data over time from media decay or errors; detected by checksums and scrubbing.
- Block storage
- Storage addressed as numbered fixed-size blocks, such as a disk, LUN or cloud volume, organised by the host's file system.
C
- Cache
- A small, fast memory or storage layer holding copies of frequently used data to hide the latency of slower layers.
- Ceph
- Open-source distributed storage system that provides block, file and object storage from clusters of commodity servers.
- Checkpoint
- A saved copy of a running job's state, such as AI model weights and optimiser state, used to resume after failures.
- CXL (Compute Express Link)
- Open interconnect on PCIe physical links that lets CPUs coherently attach extra memory and accelerators, including pooled and shared memory.
D
- DAS (direct-attached storage)
- Storage connected directly to one computer, such as an internal SSD or a USB drive, with no network in between.
- DDR5
- Fifth generation of double-data-rate SDRAM for PCs and servers, with two independent subchannels per module and on-die ECC.
- Deduplication
- Storing identical blocks or files only once and replacing repeats with references, saving capacity.
- DRAM
- Dynamic RAM: main memory that stores each bit as charge in a tiny capacitor and must be refreshed every few milliseconds.
- DWPD (drive writes per day)
- SSD endurance rating: how many times the full capacity can be written each day over the warranty period.
E
- ECC (error-correcting code)
- Extra check bits that let memory or storage detect and correct bit errors automatically.
- Erasure coding
- Splitting data into fragments plus parity fragments across devices or nodes so it survives several failures with less overhead than full copies.
F
- Fibre Channel
- Dedicated, lossless storage-networking technology used to build SANs, with 32, 64 and 128 Gb/s generations.
- File system
- Software that organises blocks into named files and directories with metadata, permissions and space allocation.
- Flash translation layer (FTL)
- SSD firmware that maps host logical addresses to physical NAND pages, hiding erase-before-write and wear.
G
- Garbage collection
- SSD background process that copies valid pages out of partly stale blocks and erases them for reuse.
- GiB (gibibyte)
- Binary unit of 2^30 bytes (1,073,741,824); a decimal gigabyte (GB) is 10^9 bytes.
H
- HAMR (heat-assisted magnetic recording)
- Hard-drive recording that briefly heats each spot with a laser so smaller, more stable magnetic grains can be written.
- HBM (High Bandwidth Memory)
- Stacked DRAM dies joined by through-silicon vias and placed beside a processor on a very wide bus, giving terabytes per second.
- HDD (hard disk drive)
- Storage device that records data magnetically on spinning platters read by heads on a moving actuator arm.
- Hyperconverged infrastructure (HCI)
- Servers that combine compute, virtualisation and software-defined storage pooled from their local drives.
I
- Immutable backup
- A backup copy that cannot be changed or deleted until its retention period ends, protecting against ransomware.
- IOPS
- Input/output operations per second: how many separate reads or writes a device or system completes each second.
- iSCSI
- Protocol carrying SCSI block commands over TCP/IP, letting ordinary Ethernet networks act as a SAN.
K
- KV cache
- Stored attention keys and values from earlier tokens that let a language model generate new tokens without recomputing them.
L
- Latency
- Time taken to complete one operation, from request to response, measured in nanoseconds, microseconds or milliseconds.
- LTO (Linear Tape-Open)
- Open magnetic-tape format family used for backup and archive, with each generation adding capacity.
- LUN (logical unit number)
- A virtual disk presented by a storage array to hosts over a SAN, identified by a number.
M
- Memory hierarchy
- The layered arrangement of registers, caches, DRAM, SSDs, HDDs and tape that trades speed against capacity and cost.
N
- NAND flash
- Non-volatile memory storing charge in cells arranged in strings; read and written in pages, erased in larger blocks.
- NAS (network-attached storage)
- A file server that shares folders over a network using NFS or SMB.
- NFS (Network File System)
- File-sharing protocol from the Unix world; current versions are NFSv3 and NFSv4.x, usually on TCP port 2049.
- Non-volatile memory
- Memory or storage that keeps data without power, such as flash, MRAM or magnetic disks.
- NUMA (non-uniform memory access)
- Design in which each CPU socket has local memory that is faster to reach than other sockets' memory.
- NVMe (Non-Volatile Memory Express)
- Storage protocol designed for flash over PCIe, allowing up to about 64K I/O queues, each up to about 64K commands deep.
- NVMe-oF (NVMe over Fabrics)
- Extension of NVMe across networks such as RDMA Ethernet, InfiniBand, TCP and Fibre Channel.
O
- Object storage
- Storage that keeps whole objects with keys and metadata in a flat namespace, accessed over HTTP APIs such as S3.
- Over-provisioning
- Spare NAND capacity hidden from the user that gives an SSD room for garbage collection and wear levelling.
P
- Page cache
- Operating-system memory holding recently read or written file data so later access avoids the disk.
- Parallel file system
- Shared file system that stripes data across many servers so thousands of clients can read and write at once, such as Lustre.
- PersistentVolumeClaim (PVC)
- A Kubernetes request for storage of a given size, access mode and class, fulfilled by a persistent volume.
Q
- QLC (quad-level cell)
- NAND that stores four bits per cell: denser and cheaper per GB than TLC but slower to write and less durable.
- Queue depth
- Number of I/O requests outstanding at once; higher depths let SSDs work in parallel but can raise latency.
R
- RAID
- Redundant Array of Independent Disks: combining drives for speed, redundancy or both, such as RAID 1, 5, 6 and 10.
- Refresh
- Periodic rewriting of DRAM rows (typically every 32-64 ms) so leaking capacitor charge does not lose data.
- Rowhammer
- Disturbance error in which repeatedly activating one DRAM row flips bits in neighbouring rows.
- RPO (recovery point objective)
- The maximum acceptable amount of data loss, measured as time since the last good copy.
- RTO (recovery time objective)
- The maximum acceptable time to restore a service after an outage.
S
- S3 API
- Amazon S3's HTTP interface for buckets and objects, now the common language of object storage.
- SAN (storage area network)
- A dedicated network giving servers shared block storage (LUNs) over Fibre Channel, iSCSI or NVMe-oF.
- SAS (Serial Attached SCSI)
- Enterprise drive interface carrying SCSI commands, supporting dual ports and expanders for many drives.
- SATA
- Serial ATA: the common consumer drive interface, limited to about 600 MB/s and, with AHCI, a single 32-command queue.
- SMART
- Self-Monitoring, Analysis and Reporting Technology: health and error counters reported by drives.
- SMB (Server Message Block)
- Windows file-sharing protocol, current dialect 3.1.1, usually on TCP port 445.
- Snapshot
- A point-in-time, usually space-efficient view of a volume or file system used for fast rollback — not a backup by itself.
- SRAM
- Static RAM built from six-transistor latches; fast and refresh-free, used for CPU caches but costly per bit.
- SSD (solid-state drive)
- Storage device built from flash memory and a controller, with no moving parts.
- Storage class (cloud)
- A price and performance tier in cloud object storage, such as S3 Standard, Standard-IA or Glacier Deep Archive.
T
- TBW (terabytes written)
- SSD endurance rating: the total amount of host data that can be written over the drive's warranty.
- Tiering
- Automatically moving data between faster, costlier storage and slower, cheaper storage according to how often it is used.
- TRIM / UNMAP
- Commands telling an SSD or array which blocks no longer hold valid data, helping garbage collection and thin provisioning.
V
- Volatile memory
- Memory that loses its contents when power is removed, such as DRAM and SRAM.
W
- Wear levelling
- Spreading writes evenly across all flash blocks so no block wears out early.
- Write amplification
- Ratio of data physically written to flash versus data the host asked to write; lower is better for speed and endurance.
Z
- ZFS (OpenZFS)
- Copy-on-write file system and volume manager with end-to-end checksums, snapshots, RAID-Z and self-healing.