SAN vs NAS vs object storage - which one to choose · Linux

Enterprise storage comes in three families: block (SAN), file (NAS), and object. Each has a different access protocol, a different scale model, and a different cost profile. Choosing the right one for the workload is a fundamental architectural decision.

The three families

Read-only / Safelsblk all— lsblk shows all block devices regardless of transport. sda is a local disk (TRAN: 'sata' or 'nvme' usually). mpath0 is a multipath device (SAN LUN presented through multiple paths). loop0 is a loop device. The transport column reveals the underlying access method.

$ lsblk -o name,size,type,tran,mountpoint,rm

NAME   SIZE TYPE  TRAN MOUNTPOINT   RM
sda    100G disk       /              0
sdb    1T   disk       /data          0
mpath0 5T  mpath      /mnt/nfs       0
loop0  100M loop       /mnt/iso       0

Family	Access protocol	Scale	Best for
Block (SAN)	Fibre Channel, iSCSI, SAS	TB to PB	Databases, VMs, low-latency workloads
File (NAS)	NFS, SMB/CIFS	TB to PB	File shares, home directories, content management
Object	S3, Swift	PB to EB	Backups, archives, media, big data

SAN characteristics

Read-only / Safeby-path— by-path shows the bus path of each disk. iSCSI paths look like ip-x.x.x.x:3260-iscsi-... and identify the SAN LUN. PCI paths identify local controllers. The kernel uses multipath to balance across redundant paths to the same SAN LUN.

$ ls -l /dev/disk/by-path/ | head -5

ip-10.0.0.10:3260-iscsi-... -> ../../sda
pci-0000:00:1f.2-ata-1 -> ../../sdb

SAN characteristics:

Block-level access - the OS sees raw devices (or multipath devices).
Low latency (sub-millisecond for FC, single-digit ms for iSCSI).
High throughput (multi-GB/s per link).
High IOPS (millions per array for modern all-flash).
Expensive per TB compared to NAS.
Mature multipath support (device-mapper-multipath).

NAS characteristics

Configuration changeNFS mount— NFS v4.1+ is the modern default. TCP transport. The mount options control performance: rsize/wsize match the server's block size, noatime reduces metadata writes, async allows the client to pipeline requests. NFS-Ganesha provides POSIX ACLs over NFS.

$ mount -t nfs -o vers=4.1,proto=tcp server:/export /mnt/nfs

NAS characteristics:

File-level access - the OS sees a filesystem.
Network-mounted - shared by many clients.
Moderate latency (single-digit ms for local network).
Moderate throughput.
Mature caching on the client side.
Lower cost per TB than SAN.

Object storage characteristics

Object storage (S3, Swift, Azure Blob) is the modern choice for unstructured data at scale:

HTTP-based access (REST API).
Effectively infinite scale (PB to EB).
Higher latency (tens of ms typical).
Eventual consistency model.
Pay-per-use pricing (cloud).
No traditional filesystem interface - tools must use the S3 API or a FUSE shim (s3fs, goofys).

Read-only / SafeS3 list— The aws s3 CLI lists objects in a bucket. Object storage uses a flat namespace (no directories) with key prefixes that look like paths. Use S3 for backups, archives, and media where the access pattern is HTTP-based and the latency budget is in tens of ms.

$ aws s3 ls s3://my-bucket/ 2>/dev/null | head

2026-01-15 10:23 backup-2026-01-15.tar.gz
2026-01-15 11:45 logs-2026-01-15.tar.gz