The path by which the recording server connects to storage determines recording continuity as much as the RAID layout. Direct-attached disk (DAS), block-level network storage (iSCSI SAN) and file-level network storage (NAS, SMB/NFS) carry different risk profiles in terms of performance, scaling, VMS support and behavior during failures. Below we compare the three options against the sustained write load of video recording and explain how to build the storage network and how to measure performance before commissioning. Which RAID level the array uses is a separate decision; we covered it in the RAID selection for video recording article.
Storage connection options for video recording
In enterprise camera systems, video recording doesn’t resemble typical IT workloads. Writes are high and sustained; reads are irregular (playback, export, archiving). Even when the bitrate drops at night, the load doesn’t fall to zero. That’s why the criterion for the storage connection isn’t peak IOPS but carrying sustained write bandwidth without latency rising. A storage proposal written for a database server usually doesn’t answer this question.
- DAS: Disks are inside the server or in a JBOD/RAID enclosure connected via SAS. There’s no network in between.
- iSCSI SAN: The storage array presents a block device (LUN) to the server over Ethernet. The server formats this volume like a local disk.
- NAS (SMB/NFS): The storage device manages the file system; the recording server writes files to a network share.
DAS, iSCSI and NAS comparison
| Criterion | DAS | iSCSI SAN | NAS (SMB/NFS) |
|---|---|---|---|
| Access level | Block, local | Block, over the network | File, over the network |
| VMS support | Universal | Common (appears as a local disk) | Varies by manufacturer |
| Network dependency | None | High; a separate network is needed | High |
| Scaling | Limited by the server chassis/JBOD | Array expansion, multiple servers | Easy to add capacity |
| Sharing with failover | Difficult | Possible (LUN is reassigned) | Share is reconnected |
| Point of failure | Server and controller | Network, array controller | Network, NAS head |
| Suitable scenario | Single server, medium scale | Multiple servers, central storage | Archive tier, small installations |
DAS gives the most predictable performance with the fewest components. In installations where storage per server is fixed, it’s still the most robust choice. Its weak side is capacity and failover flexibility: when the server goes, the disks become inaccessible with it. iSCSI lets several recording servers share a central array and redistribute capacity between servers; in return, network design becomes part of storage. A network fault is now a storage fault. NAS is the easiest option to set up; but the overhead of the file protocol, locking behavior and whether the VMS supports live writing to a network share decide the matter.
VMS support and file systems
Read the VMS manufacturer’s supported storage document before choosing. These are situations we encounter on site again and again:
- Live recording vs. archive: Some VMSs support only local or block storage for live recording while allowing a network share (UNC path) in the archive tier. In that case, the NAS’s natural place is the archive tier; the tiered structure is explained in the video recording archiving and backup article.
- Service account and authentication: A VMS service writing to an SMB share should run not as the local system account but as a service account with permissions on the share. A service account whose password expires is one of the most common reasons recording silently stops. Nobody notices, until footage is requested.
- File system and allocation unit size: An iSCSI LUN is formatted on the server side. On Windows, check the allocation unit size the VMS recommends for NTFS or ReFS. A large allocation unit is usually suitable for large video files, but choose the value according to the manufacturer’s recommendation; changing it after formatting requires moving the data.
- The single-writer rule: If you connect an iSCSI LUN to two servers at once without a cluster file system, the file system gets corrupted. In a VMS failover scenario, make sure the LUN is connected only to the active server.
Storage network design
If iSCSI or NAS is chosen, storage traffic is separated from camera traffic. The typical setup on a recording server: interfaces facing the camera network and interfaces facing the storage network are physically separate. The camera stream enters the server from one network, and the same data leaves for storage over a second network. On a full-duplex link, incoming camera traffic and outgoing storage writes use different directions; the real problems are that data read from storage during playback and archiving shares the same direction as camera traffic, that switch buffers are shared, and that a broadcast storm or loop on the camera network also cuts storage access. A loop takes down two systems at once.
- A separate VLAN and preferably separate switches: The storage network should run in a non-routed VLAN with no gateway. In large installations, a separate pair of switches is used for storage.
- Link speed: Calculate the sustained write bandwidth per recording server and leave at least double the margin; read traffic during archiving, playback and rebuild also uses the same link. The calculation is detailed in the camera network bandwidth calculation article.
- Switch buffers: In setups where many servers write to a single storage port, access switches with shallow buffers drop packets under burst load. For the storage network, prefer data-center-class switches with documented buffer capacity. Rising output drops in interface counters are the first clue.
- Jumbo frames: If MTU 9000 is used, the server interface, switch ports and storage interface must be configured with the same value. Verify the path with the test below; 8972 bytes is the 9000-byte MTU minus the IP and ICMP headers (28 bytes).
# Linux: don't-fragment, 9000 MTU path test
ping -M do -s 8972 -c 4 10.20.0.10
# Windows: don't-fragment (-f), payload size (-l)
ping -f -l 8972 10.20.0.10
If a standard-size ping works but this one doesn’t, there’s an MTU mismatch on the path. Either fix it end to end or don’t use jumbo frames at all. A half-done configuration gives worse results than standard MTU: small packets pass, large writes stall and the fault appears intermittent. An intermittent fault is the most expensive kind.
Multipathing and redundancy
In iSCSI, the right method for both redundancy and bandwidth is multipath I/O. Each server interface gets its own IP address, each controller of the storage array has separate target portals, and the operating system combines the paths to the same LUN. Link aggregation with LACP isn’t the recommended method for iSCSI; a single iSCSI session usually uses a single physical link, and path failure detection is done more controllably in the operating system’s MPIO layer. The protocol itself is defined by RFC 7143.
On a Windows Server-based recording server, the basic configuration looks like this (addresses are examples):
# Start the iSCSI initiator service automatically
Set-Service -Name MSiSCSI -StartupType Automatic
Start-Service -Name MSiSCSI
# Install the MPIO feature (a restart may be required)
Install-WindowsFeature -Name Multipath-IO
# Automatically claim iSCSI disks for MPIO
Enable-MSDSMAutomaticClaim -BusType iSCSI
# Default load balancing policy: Round Robin
Set-MSDSMGlobalDefaultLoadBalancePolicy -Policy RR
# Add the target portal of both controllers and connect over two paths
New-IscsiTargetPortal -TargetPortalAddress 10.20.0.10 -InitiatorPortalAddress 10.20.0.21
New-IscsiTargetPortal -TargetPortalAddress 10.20.1.10 -InitiatorPortalAddress 10.20.1.21
# If the array presents more than one target, choose the right one
$t = Get-IscsiTarget | Select-Object -First 1
Connect-IscsiTarget -NodeAddress $t.NodeAddress -IsMultipathEnabled $true -IsPersistent $true `
-InitiatorPortalAddress 10.20.0.21 -TargetPortalAddress 10.20.0.10
Connect-IscsiTarget -NodeAddress $t.NodeAddress -IsMultipathEnabled $true -IsPersistent $true `
-InitiatorPortalAddress 10.20.1.21 -TargetPortalAddress 10.20.1.10
On Linux-based recording servers, open-iscsi and dm-multipath are used:
iscsiadm -m discovery -t sendtargets -p 10.20.0.10
iscsiadm -m discovery -t sendtargets -p 10.20.1.10
iscsiadm -m node --login
multipath -ll
In the multipath -ll output, check that the expected number of paths for each LUN shows as “active.” After installation, don’t close this step without pulling a cable and seeing that the path drops and recording continues. If the storage manufacturer recommends a specific load balancing policy (for example the preferred controller path on ALUA-capable arrays), that recommendation is applied.
To complete redundancy, the two paths must pass through two different switches, and the storage array must have dual controllers. On a single-controller array, multipathing only protects against cable and switch failures. When the controller goes, everything goes. Considered together with recording server failover, this setup is the storage leg of VMS high availability design. Shared storage options in Hyper-V environments are also compared in the shared storage for Hyper-V article.
Performance validation
Before commissioning storage, measure sustained write capacity with a test that mimics the VMS load. Because video recording produces many simultaneous sequential write streams, the test is also set up as parallel sequential writes. A random 4K test is the wrong question for this job.
# Linux (fio): 8 parallel sequential writes, 1 MiB blocks, cache bypassed
# (each job uses a 20 GiB file; 160 GiB of free space is needed in total)
fio --name=video --directory=/mnt/recording --rw=write --bs=1M --size=20G \
--numjobs=8 --iodepth=4 --direct=1 --ioengine=libaio \
--time_based --runtime=600 --group_reporting
# Windows (DiskSpd): 100% write, sequential, 8 threads, no cache
diskspd.exe -c20G -b1M -w100 -si -t8 -o4 -d600 -Sh -L E:\test.dat
Three values are looked at in the result: average write bandwidth, latency (especially at high percentiles) and how much bandwidth fluctuates during the test. The goal is to sustain a value clearly above the calculated recording load without fluctuation. Don’t be fooled by high values in the first minutes; every array is fast until the controller cache fills. Repeat the test in a degraded state too, with a disk removed or while a rebuild is running; that’s the performance on your worst day. On the live system, feed recording server warnings such as “disk queue,” “dropping frames” or “recording buffer full” into the monitoring system.
Selection checklist
- The storage types the VMS manufacturer supports for live recording and archiving have been documented.
- Sustained write bandwidth per recording server has been calculated, and margin has been left in link capacity.
- Storage traffic is carried on a separate VLAN and separate physical interfaces.
- If jumbo frames are used, they’ve been verified end to end; if not, this has been noted as a deliberate decision.
- MPIO/dm-multipath has been configured for iSCSI, and the two paths pass through two different switches.
- The storage array has dual controllers, or the single-controller risk has been accepted and documented.
- Password and permission management has been defined for the service account using the SMB share.
- Write tests in normal and degraded states have been run with fio or DiskSpd, and the results kept.
- Storage latency and recording warnings have been connected to the monitoring system.
To plan the storage architecture in terms of capacity and connectivity together, see the camera recording storage and capacity planning service page.
Frequently asked questions
Is it right to use a NAS for video recording?
It can be used if the VMS manufacturer supports an SMB or NFS share as a live recording target and the network and NAS can handle the sustained write load. In many installations, the NAS’s proper place isn’t live recording but the archive tier. Check the manufacturer’s supported storage list before deciding.
Can iSCSI traffic be carried on the same network as the camera VLAN?
Technically possible, not recommended. When camera and storage traffic share the same link, storage latency rises at busy moments and recording can be interrupted. Carry storage traffic in a separate VLAN, preferably over separate physical ports or switches.
Are jumbo frames mandatory?
No. Jumbo frames can reduce packet count and CPU load, but if the whole path doesn’t have the same MTU, they cause connection problems. Measure with standard MTU first; use jumbo frames only after configuring and testing them end to end.
Is LACP the same as iSCSI multipathing?
No. LACP combines several links into one logical link, and a single iSCSI session mostly uses a single physical link. The correct method for iSCSI is a setup where each path gets its own IP address and the operating system manages paths with MPIO or dm-multipath.