If a backup job that started at night is still running at 09:00, the problem is usually not Veeam but the architecture: roles piled onto a single server, a proxy that fell back to network mode, or a repository with more tasks than it can handle. In a small environment, installing Veeam Backup & Replication on a single server and leaving all roles there works. When the environment grows to hundreds of virtual machines, multiple storage systems and a second site, however, the backup window, restore speed and resilience come to depend on where and with what capacity the components are placed. Below we go through the components, transport modes, Hyper-V-specific proxy options, sizing, the Scale-Out Backup Repository structure and reading bottlenecks in order.
Veeam components and their functions
| Component | Function | Placement note |
|---|---|---|
| Backup server | Schedules jobs, manages components, uses the configuration database | Also takes on the default proxy and repository roles at installation; stays out of the data path when roles are separated |
| Backup proxy | Reads data from the source, compresses, deduplicates and sends it to the repository | The point closest to the source storage |
| Backup repository | Stores backup files; Windows, Linux, SMB, deduplication appliance or object storage | Fast disk with sufficient capacity |
| Gateway server | Processes data on behalf of the repository for SMB shares and some appliances | Close to the repository |
| Mount server | Mounts backups for file-level restore and instant recovery | Usually the repository server |
| Guest interaction proxy | Connects to the guest operating system for application-aware processing and indexing | A point with network access to the VMs |
| WAN accelerator | Reduces WAN traffic for inter-site backup copies and replication | One at the source and one at the target site |
The configuration database can run on Microsoft SQL Server or PostgreSQL. When the backup server is lost, backups aren’t lost but jobs stop. Take Veeam’s built-in configuration backup, encrypted, to a repository different from the main one; in a configuration backup taken without encryption, stored credentials don’t come back and you have to re-enter the password of every component. Whether the backup server should be a member of the production domain is a separate security design decision; we don’t go into it here.
Comparison of transport modes
In a VMware vSphere environment, the transport mode determines how the proxy reads VM disks. In automatic selection, Veeam first tries direct storage access, then virtual appliance mode, and finally network mode. The order is right; the result isn’t always.
| Mode | How it reads | Advantage | Limitation |
|---|---|---|---|
| Direct storage access (Direct SAN / Direct NFS) | The proxy accesses datastore LUNs or the NFS share directly | Puts no load on ESXi hosts, high speed | Requires a physical proxy and a storage network connection; restore with Direct SAN only for thick disks |
| Virtual appliance (hot-add) | The proxy, which is a VM, attaches the disks of the VMs being backed up to itself | Needs no additional hardware, good speed | The proxy must be able to access the same datastores; adds attach/detach time |
| Network (NBD/NBDSSL) | Reads through the ESXi management interface | Works in every environment, simple setup | The management network and the ESXi side can become a bottleneck; slow for large full backups |
Rule of thumb: Direct SAN with a physical proxy in server and storage infrastructures with large block storage and access to the Fibre Channel/iSCSI network; hot-add proxy VMs in every cluster in fully virtualized, small to mid-sized environments; network mode in remote branches and single-host environments. When backups are slow, this is the first place we look in the console: automatic selection may have silently fallen back to network mode. Typical causes are a hot-add proxy not seeing the VM’s datastore, or LUNs not being presented to the proxy for Direct SAN. Check the mode used in the job session log from the “[san]”, “[hotadd]” or “[nbd]” tags.
On-host and off-host proxy in a Hyper-V environment
In Hyper-V there is no separate transport mode choice. There is one decision: who will read the data?
- On-host backup: The Hyper-V host itself takes on the source proxy role. It is simple to set up and needs no additional hardware. But compression and deduplication use the host’s CPU and memory; during the backup window it competes with production VMs for the same resources.
- Off-host backup: A separate physical Windows Server with the Hyper-V role installed acts as the proxy. The storage system’s hardware VSS provider must offer transportable shadow copy support; the snapshot is attached to the proxy and the data is read from there. The off-host scenario is also possible with SMB3 shared storage. Check the vendor’s current requirements for compatibility of the proxy’s operating system version with the Hyper-V hosts.
If there is noticeable CPU pressure on production hosts during the backup window, or the backup job spills into working hours, consider an off-host proxy. The point where we get stuck most often in off-host setups is the storage vendor’s hardware VSS provider not being installed on the proxy, or its version not matching the one on the hosts; in that case the job fails at the snapshot stage. Shared storage options and VSS provider compatibility are covered in the choosing shared storage for Hyper-V article.
Sizing the proxy and repository
In Veeam, a task is the processing of one VM disk. A VM with four disks is processed in parallel with four tasks if the concurrent task limit allows. Both the proxy and the repository have a “maximum concurrent tasks” setting, and resources are sized according to this number. Since the cores and memory needed per task change from version to version, take them from the system requirements document of the version you use. Don’t size from memory.
Worked example
Assumptions: 60 TB of used source data, a 48-hour weekend window for the weekly active full backup, a transfer rate of 100 MB/s per task measured in a pilot test, 1 core and 2 GB RAM per task (example values; verify with the version document).
- Total speed required: 60,000,000 MB ÷ 172,800 s ≈ 347 MB/s.
- Minimum concurrent tasks required: 347 ÷ 100 ≈ 4 tasks.
- Leaving headroom for the restore scenario, several systems coming back at once and job retries, the design target is 8–12 tasks.
- This corresponds to 6 cores and 12 GB RAM on each of two proxies, excluding operating system resources. The second proxy is also redundancy: a single proxy is a single point of failure.
On the repository side, limit the number of concurrent tasks according to the disk’s random write capacity. If the proxy sends data with 12 tasks while the repository can handle 4, the bottleneck just moves. The problem isn’t solved, it’s relocated. Repository capacity itself is a separate calculation and is described in the backup storage capacity calculation article. In new repositories, use per-machine backup files; parallel writes get faster, and moving and deleting a single VM becomes easier.
Scale-Out Backup Repository
A Scale-Out Backup Repository (SOBR) presents multiple repositories as a single logical target and divides them into tiers. When a repository fills up, you add an extent instead of moving jobs one by one to a new target.
- Performance tier: Extents on block storage. The placement policy offers two options: Data locality keeps all files of a chain on the same extent and is compatible with ReFS/XFS block cloning; Performance distributes full and incremental files across different extents, but losing one extent affects the whole chain.
- Capacity tier: Object storage such as S3-compatible or Azure Blob. In copy mode, each backup is also written to object storage as soon as it is created; in move mode, sealed chain files older than a set period are moved and local space is freed.
- Archive tier: Low-cost tiers with long retrieval times, such as Amazon S3 Glacier or Azure Archive, for long-term GFS points.
If an extent is to be taken into maintenance, new writes are stopped with “maintenance mode,” and if needed the backups on it are moved to other extents with “evacuate.” Ransomware goes for the backup console and the repository first. That’s why you should turn on immutability in the object storage tier from the start; the options are compared in the immutable backup options article.
Reading the bottleneck in task statistics
In each job session’s statistics, Veeam shows how busy the four stages of the data path are. Read this line before guessing:
Load: Source 92% > Proxy 35% > Network 18% > Target 9%
Primary bottleneck: Source
| Bottleneck | Likely cause | What to do |
|---|---|---|
| Source | Read capacity of production storage, network mode | Check the transport mode, monitor storage latency |
| Proxy | Insufficient CPU, high compression level | Add cores, balance the task count |
| Network | 1 GbE link, WAN, wrong interface | Separate the data network, increase bandwidth |
| Target | Repository disks, merge operations, parity write penalty | Review the repository disk layout and the concurrent task limit |
Every data path has a bottleneck. This is normal. The real question is whether the bottleneck pushes past the backup window or the restore time. If the Source ratio is high but the job duration is within target, leave it alone. On incremental backups, however, Source appearing high together with a low total speed is a separate alarm: CBT/RCT may not be working and the whole disk may be read from the start. This mostly appears after a VM migration or a host update.
Design checklist
- Is the backup server outside the data path, on a separate, protected server? Is the configuration backup taken, encrypted, to a different repository?
- Are there at least two proxies in each cluster? Has the transport mode used been verified from the job logs?
- In Hyper-V, has the impact of the on-host proxy on production hosts been measured?
- Are the proxy and repository concurrent task counts consistent with each other?
- Does the repository formatting support block cloning, and is the SOBR placement policy suited to it?
- Does the second copy go to a different location, and at least one copy to an immutable repository?
- Are bottleneck statistics reviewed weekly?
For small environments, the starting point is the free backup with Veeam Community Edition article; for designing replication to a second site, see the Hyper-V Replica and Veeam replication article. You can find how we design and build Veeam architecture on our backup and disaster recovery and virtualization and data management pages. For current component requirements, use the Veeam Help Center, and for design recommendations, Veeam Best Practices.
Frequently asked questions
Can the Veeam proxy and repository be on the same server?
They can; this is common in small and mid-sized environments and reduces network traffic. Since resources are shared, size the concurrent task counts together. In large environments, placing proxies close to the source storage and the repository on separate servers close to the storage makes scaling easier.
Which transport mode should be chosen in a VMware environment?
Direct SAN if there is a physical proxy and SAN access, hot-add proxy VMs in fully virtual environments, and network mode for remote branches and small environments. Verify which mode was used in automatic selection from the job session log.
When is an off-host proxy needed in Hyper-V?
An off-host proxy is considered if noticeable CPU and memory pressure occurs on production hosts during the backup window, or the window isn’t enough as the environment grows. The storage system must provide a hardware VSS provider that supports transportable shadow copies.
What does it mean if Source is high in the bottleneck statistics?
Reading data from production storage is the slowest stage. Check the transport mode, storage latency and whether changed block tracking works for incremental backups. If the job duration is within target, no action is needed.