Skip to content

Hyperconverged infrastructure (HCI) from data centres in Germany: clusters from 3 nodes, whose local NVMe form a shared pool — the storage is not located next to the cluster, but inside it. Billed per node and month, the licence per core; how much of the pool is usable is determined not by the price, but by the number of nodes.

PC-S32 is not approved because the class does not have a separate storage fabric.

Per node from
€1,690.20$1,960.63
PC-M64 incl. licence, net
Minimum cluster
3 nodes
€5,109.60$5,927.14 net, storage and private network included
Usable capacity
47% – 76%
depending on the protection method, 3 to 24 nodes
Licence
€13.90$16.12
per core and month

Class, number of nodes, location — the pool, the usable share and the total are updated next to it after each decision; what you select here is used as a preset by the configurator.

Node classAmounts per node, with licence

48 Genoa cores with 12 memory channels — the most frequently booked production class.

Number of nodes

5.41 TB47% usable

3 nodes24 nodes

With 3 nodes, the pool runs with Two-way mirroring (RF2): usable 5.41 TB of 11.52 TB raw, tolerates one failure. From 6 nodes, Erasure coding 4+2 applies, usable 62%.

Mirror 1+1one block per nodeN1D1N2D1N3usable 47% · 5.41 TBraw 11.52 TB
  • Data block
  • Mirror copy
  • Blocks of other stripes
  • Not usable — Parity or mirror, metadata, reserve
  • Free space — up to 24 nodes

Location

The locations offered are those where PC-M64 is located.

Recovery to a second region

Without a second region, the pool protects against node failures in the cluster, not against the failure of the location.

Term

− €255.48$296.36Discount / month

The discount applies to the monthly net amount. With a term, you pay for the entire period in advance with the first invoice.

Net per month after discount€4,854.12$5,630.78€5,776.40$6,700.63 incl. VATSummary

The storage pool is created from the local NVMe of all nodes — the flash drives plugged directly into the node. What remains for guest data is determined not by price, but by the protection method. Which one applies depends on the number of nodes.

Below 6 nodes, the pool runs two-way mirrored: every block is written twice, meaning half the raw capacity is allocated before a single guest byte is stored. An additional 3.0% goes to metadata and the reserve used to rewrite data after a node failure — leaving 47% usable.

From 6 nodes, erasure coding takes effect. Instead of a full copy, there are only 2 parity blocks alongside 4 data blocks. This is the most valuable leap on this page: double the number of nodes, 2.64 times the usable capacity. Every subsequent step yields less — from 18 nodes onwards, it is only 5 percentage points.

The tiers are seamless: every permitted cluster size between 3 and 24 nodes falls into exactly one protection method. There is no size where you have to check what applies first.

Hyperconverged

47% to 76% · baseline at zero

How much raw capacity remains usable

Share of raw capacity0%25%50%75%100%47%53%62%38%71%29%76%24%Erasure coding from 6 hosts3 hosts+15 % pts+9 % pts+5 % pts3–5 hostsDoubleMirroring (RF2)1 host failure6–11 hostsErasure coding 4+22 host failures12–17 hostsErasure coding 6+22 host failures18–24 hostsErasure coding 8+22 host failuresENTRONYX HCI: 3 to 24 hosts per cluster
  • Two-way mirroring (RF2) — usable 47%, one host failure
  • Erasure coding — usable 62% to 76%, two host failures
  • Not usable — Parity or mirror, metadata, recovery reserve

The full column height is the raw capacity of all local NVMe in the cluster. The capacity jumps, it does not grow: there is no intermediate value between two steps, hence rectangles and not a line. The gain at each edge becomes smaller — from 3 to 6 hosts it is 15 percentage points, from 12 to 18 only 5. From then on, expansion is worthwhile for computing power, no longer for storage.

How the usable share is calculated

The raw proportion follows solely from the stripe geometry: data blocks divided by data blocks plus redundancy. What is deducted after that is the reserve for metadata and for the space to which data is rewritten after a node failure.

  • Two-way mirroring (RF2)1 node failure
    Nodes
    3–5
    Stripe (blocks)
    1+1 (2)
    Raw
    50.0%
    Reserve
    − 3.0%
    Usable
    47%
    Raw capacity
    11.52 TB
    Usable capacity
    5.41 TB
  • Erasure coding 4+22 node failures
    Nodes
    6–11
    Stripe (blocks)
    4+2 (6)
    Raw
    66.7%
    Reserve
    − 4.7%
    Usable
    62%
    Raw capacity
    23.04 TB
    Usable capacity
    14.28 TB
  • Erasure coding 6+22 node failures
    Nodes
    12–17
    Stripe (blocks)
    6+2 (8)
    Raw
    75.0%
    Reserve
    − 4.0%
    Usable
    71%
    Raw capacity
    46.08 TB
    Usable capacity
    32.72 TB
  • Erasure coding 8+22 node failures
    Nodes
    18–24
    Stripe (blocks)
    8+2 (10)
    Raw
    80.0%
    Reserve
    − 4.0%
    Usable
    76%
    Raw capacity
    69.12 TB
    Usable capacity
    52.53 TB

In numbers for the entry-level class: A cluster of 3 × PC-M64 carries 11.52 TB raw and 5.41 TB usable, the same cluster with 6 nodes 14.28 TB. The two right columns apply to PC-M64 with 3.84 TB local NVMe per node, calculated at the lower limit of the respective tier. Within a tier, the capacity continues to grow with each node — the usable proportion remains the same until the next tier takes effect.

The licence surcharge is the price of the storage

The price of a managed node is the bare metal plus platform operations. The cluster licence is added per core and month — and with it the entire primary storage, for which a separate datastore is purchased under vSphere.

The licence costs €13.90$16.12 per core and month, with a minimum purchase of 16 cores per socket. On PC-M64, that is 48 cores and thus €667.20$773.95 per node and month.The minimum purchase does not apply to any approved class — the smallest comes with 48 cores per socket. Compared to vSphere, that is €4.00$4.64 more per core, so €576.00$668.16 per month in the minimum cluster. In return, an entire item is eliminated, the datastore: the local NVMe is included in the bare metal price in both cases, hyperconvergence simply combines them into a shared pool.

The minimum cluster comprises 3 nodes, not two. A pool of two nodes cannot decide which half is allowed to continue writing after a network split — both would see themselves as the survivor. Only the third node establishes the majority. At the upper end, the cluster is limited to 24 nodes.

  • Distributed storage pool across the local NVMe of all hosts
  • Data locality: the active block resides on the VM's host
  • Erasure coding or mirroring depending on cluster size
  • Inline deduplication and compression, can be disabled per container

Minimum cluster, item by item

3 × PC-M64 — the smallest configuration in which the pool can form a quorum, i.e. a majority that decides after a failure. A datastore is not on the list because the storage is in the nodes; the private network, however, is: under hyperconvergence, the configurator does not allow a level smaller than vRack 10 Gbit/s.

3 × PC-M64144 cores · 696 GB allocatable
€3,069.00$3,560.04
ENTRONYX HCI — licence€13.90$16.12 per core and month
€2,001.60$2,321.86
Storage pool in the cluster5.41 TB usable · Two-way mirroring (RF2)
included in the cluster
vRack 10 Gbit/sper cluster
€39.00$45.24
1 Gbit/s guaranteed
in the node price
Platform operations
in the node price
Net total per month
€5,109.60$5,927.14
Gross per monthincl. 19% VAT.
€6,080.42$7,053.30

The same cluster under vSphere

3 × PC-M64, same number of nodes, same usable capacity — here via a booked Datastore NVMe Performance of 5,414 GB, the performance class that the pool delivers from the local NVMe.

3 × PC-M64144 cores · 696 GB allocatable
€3,069.00$3,560.04
ENTRONYX Managed vSphere — licence€9.90$11.48 per core and month
€1,425.60$1,653.70
Datastore NVMe Performance5,414 GB
€590.13$684.55
vRack 1 Gbit/sper cluster
€0.00$0.00
1 Gbit/s guaranteed
in the node price
Platform operations
in the node price
Net total per month
€5,084.73$5,898.29
Hyperconvergence
more expensive by €24.87$28.85

The surcharge of €24.87$28.85 per month buys Storage in the performance class of local NVMe, one procurement instead of two, data locality for reads and from 6 nodes two tolerated node failures.

Against Datastore NVMe Standard (0.9 ms latency, NFS 4.1 or iSCSI), the same cluster would be €4,884.41$5,665.92; Hyperconvergence would then be €225.19$261.22 more expensive — it replaces the fast class, not the slow one.

The calculation applies to this one capacity. If storage requirements grow faster than compute requirements, it shifts in favour of the datastore: it can be expanded individually, the pool only via additional nodes.

HCI cluster versus cluster with external storage

Both stacks run on the same node classes. The difference lies in where the primary storage is located — and almost everything else follows from this: the minimum size, the maintenance duration, the recovery levels and the question of whether capacity can grow individually.

  • Location of primary storage
    ENTRONYX HCI
    Local NVMe of all nodes, combined into a pool
    ENTRONYX Managed vSphere
    External datastore, booked separately
  • Licence rate
    ENTRONYX HCI
    €13.90$16.12
    ENTRONYX Managed vSphere
    €9.90$11.48
  • Licence per PC-M641 sockets, 48 cores
    ENTRONYX HCI
    €667.20$773.95
    ENTRONYX Managed vSphere
    €475.20$551.23
  • Cluster from … to
    ENTRONYX HCI
    3 – 24 nodes
    ENTRONYX Managed vSphere
    2 – 32 hosts
  • Approved classes
    ENTRONYX HCI
    4
    ENTRONYX Managed vSphere
    4
  • External datastore required for operationswithout it, nothing restarts automatically under vSphere
    ENTRONYX HCI
    ENTRONYX Managed vSphere
  • Protection method follows cluster sizeMirroring or erasure coding, depending on the number of nodes
    ENTRONYX HCI
    ENTRONYX Managed vSphere
  • Capacity expandable independently of the number of nodes
    ENTRONYX HCI
    ENTRONYX Managed vSphere
  • Management plane
    ENTRONYX HCI
    Web console in the cluster, REST API v3, Prometheus endpoint
    ENTRONYX Managed vSphere
    Dedicated vCenter per customer, API access, SSO against your directory
  • Maintenance windows
    ENTRONYX HCI
    Tue 02:00–06:00 CET, rolling with data redistribution before each restart
    ENTRONYX Managed vSphere
    Wed 02:00–06:00 CET, rolling host by host, announcement 14 days in advance

Where this setup is the right fit

  • One environment instead of two procurements

    Computing power and storage come from the same order, grow at the same pace and are managed in the same console. If you do not have a storage department, you do not just save money here, but an entire operations discipline.

  • Uniform guest workloads in large numbers

    Virtual workspaces, application servers, test environments: many guests with a similar profile are distributed cleanly across the nodes and keep data locality high. The pool then works predominantly locally.

  • Expansion in equal steps

    An additional node brings cores, memory and capacity in a fixed ratio. This is a limitation and at the same time the reason why capacity planning here is a number instead of a table.

  • Where hyperconvergence is the wrong fit

    Cross-check

    If storage and computing power grow at different speeds, each expansion pays for the other half as well. If you need a lot of capacity with little computing power — file services, archives, media inventories —, you are better off with dedicated hosts and a separate datastore.

Each node provides compute, memory and local NVMe in a fixed ratio. This is the condition of the design: if you need capacity, you choose either a denser class or more nodes — there is no third option.

Cluster node without cover, from above: at the front the bay with 24 drive trays, behind it the fan wall, two heat sinks and 24 memory modules.
Storage at the front, compute at the back: the class determines both at once — cores, memory and the node's contribution to the pool.
  • PC-M64PC-M
    Processor
    EPYC 9454P2.75 – 3.80 GHz · 1 sockets
    Cores / threads
    48 / 96
    Allocatable RAM
    232 GB
    NVMe per node
    3.84 TB
    Connection
    1 Gbit/s
    Usable with 3 nodes
    5.41 TB
    Licence / month
    €667.20$773.95
    Total / month
    €1,690.20$1,960.63
  • PC-M64iPC-M
    Processor
    Xeon 6741P2.50 – 3.80 GHz · 1 sockets
    Cores / threads
    48 / 96
    Allocatable RAM
    232 GB
    NVMe per node
    7.68 TB
    Connection
    1 Gbit/s
    Usable with 3 nodes
    10.83 TB
    Licence / month
    €667.20$773.95
    Total / month
    €1,760.20$2,041.83
  • PC-L128PC-L
    Processor
    EPYC 95543.10 – 3.75 GHz · 1 sockets
    Cores / threads
    64 / 128
    Allocatable RAM
    472 GB
    NVMe per node
    7.68 TB
    Connection
    5 Gbit/s
    Usable with 3 nodes
    10.83 TB
    Licence / month
    €889.60$1,031.94
    Total / month
    €2,788.25$3,234.37
  • PC-X256PC-X
    Processor
    2× EPYC 99652.25 – 3.70 GHz · 2 sockets
    Cores / threads
    384 / 768
    Allocatable RAM
    960 GB
    NVMe per node
    3.84 TB
    Connection
    5 Gbit/s
    Usable with 3 nodes
    5.41 TB
    Licence / month
    €5,337.60$6,191.62
    Total / month
    €11,130.78$12,911.70

The “Usable” column calculates using the minimum cluster method, two-way mirroring; more capacity is provided by more nodes — 6 × PC-M64 reach 14.28 TB instead of 5.41 TB — or a denser class like PC-L128 with 7.68 TB per node.

vCPU limit per class

PC-M64
288 vCPU · 6:1
PC-M64i
288 vCPU · 6:1
PC-L128
256 vCPU · 4:1
PC-X256
1.536 vCPU · 4:1

Allocatable, not recommended. In the cluster, a second reserve is added: the guests of a failed node must fit into the remaining space, and the pool needs compute time on the same nodes during recovery.

Largest permitted cluster

Nodes
24
Node class
PC-X256
allocatable vCPU
36,864
Memory allocatable
22.5 TB
Usable capacity
70.04 TB
Calculated guests
15,360

The usable capacity calculates with the broadest protection method provided in the catalogue. Larger clusters are not created by more nodes, but by multiple clusters side by side.

What restricts each class in the cluster

A class without a stated limit is a class whose limit no one has looked for yet. Here it stands for each of the 4.

  • PC-M64

    PC-M

    One socket, one NUMA node: VMs over 232 GB RAM no longer fit on this class. For wider single VMs, PC-L128 or PC-X256 is the right fit.

  • PC-M64i

    PC-M

    Live migration is only guaranteed between hosts of the same CPU generation. A mixed cluster of PC-M64 (AMD) and PC-M64i (Intel) does not migrate during operations.

  • PC-L128

    PC-L

    The boost ends at 3.75 GHz — the class wins on width, not on clock speed. Loads that depend on single threads run faster on PC-S32 with 5.7 GHz.

  • PC-X256

    PC-X

    Two sockets mean two NUMA nodes. VMs over 480 GB RAM or over 192 vCPU exceed one node and lose around 12% memory bandwidth — with licences per core, this host also pays for 384 cores.

  • PC-S32

    Not approved

    3 Gbit uplink without a separate storage fabric. Shared NFS or iSCSI datastores are not permitted on this class so that replication and guest traffic do not share the same line. Without a shared datastore, there is no automatic HA restart.

Locations for the entry level

A cluster is located entirely at one location — the pool needs runtimes in the range of a few microseconds between its nodes. Distribution across two regions is only created via the recovery levels, not via the cluster itself. PC-M64 is available at 2 locations, PC-X256 at 2.

Available for PC-M64 2

  • fra1Frankfurt am Main€1,690.20$1,960.63
  • dus1Düsseldorf€1,690.20$1,960.63
Stack of six identical nodes in an open rack on the left edge of a bright, empty hall; all six status lights are green.

From 6 nodes, Erasure coding 4+2 replaces mirroring — 62% instead of 47% of the raw capacity then becomes usable.

Reads are local, writes are across the network

The pool is not located next to the nodes, but inside them. This keeps read operations short and makes write operations a matter of the network — and it explains why the storage fabric of a class is an admission requirement here and not an equipment option.

The number of nodes involved grows with the stripe width: with mirroring 1+1, a write operation touches 2 nodes, with Erasure coding 8+2 it is 10. Wider stripes save capacity and cost network paths. Read operations remain unaffected — they hit the local block.

Write operations, data redistribution and, after a failure, recovery run over the storage fabric, separated from guest traffic at the node's connection. If it is missing, all this traffic shares a line with the guests. That is why PC-S32 is not approved here.

Write path

Where a write operation goes

Mirroring 1+1: a partner

from 3 nodes

K1GuestReadDataWriting over the networkN2CopyN3uninvolved

Erasure coding 4+2: five additional nodes

from 6 nodes

K1GuestReadDataWriting over the networkN2DataN3DataN4DataN5ParityN6Parity
The smallest cluster of each level is shown. Each block of a stripe is located on a different node — any node not in the stripe remains uninvolved during writing.
Rear view of three identical nodes in the rack, each with two optical fibres plugged in and bundled upwards to the fabric.
The fabric carries the storage traffic of the pool: 10 Gbit/s per node on PC-M64, 25 Gbit/s on PC-X256.
  • Data locality on read

    The active block of a VM resides on the node where the VM is running. A read operation thus stays within the device and does not need the network. If the VM migrates to another node, its blocks follow in the background — until then, it reads over the network and is therefore slower.

    Read path
    local NVMe, no network path
    After a migration
    blocks follow, reading remains remote in the meantime
    Cold start of a VM
    initial reads remote until the copy is ready
    Visible in
    Prometheus endpoint of the cluster, metric per container
  • The write path always goes over the network

    A write operation is only considered acknowledged when redundancy is established. With two-way mirroring, this is a partner node; with erasure coding, it is all nodes of a stripe. This is exactly where the higher usable capacity is paid for: not in euros, but in network paths per write operation.

    Mirroring
    one copy on a partner node
    Erasure coding
    data blocks and parity across the entire stripe
    Acknowledgement
    only after full redundancy, never before
    Consequence
    small write operations cost proportionally more than large ones
  • Deduplication and compression before writing

    Both run inline, i.e. before storing, and both cost compute time on the same node that serves the guests. Therefore, it can be disabled per container: for databases that already write compressed data, the gain is zero and the overhead is not zero.

    Point of effect
    before writing, not as a downstream run
    Costs
    compute time on the node that also serves guests
    Can be disabled
    per container, not just for the entire cluster
    Sensible to disable
    for already compressed or encrypted data
  • A dedicated network for storage

    Guest traffic runs via the node's connection, while the pool's storage traffic runs over a separate fabric. This is the fundamental condition of the architecture: if you bring storage into the compute node, you also bring its traffic there. Classes without a separate fabric are therefore not approved — not as a precaution, but because the commitment could otherwise not be kept.

    Separation
    Storage traffic on the fabric, guest and management traffic in dedicated VLANs
    Priority
    Guest write operations before recovery, otherwise the acknowledgement breaks off
    Recovery
    runs over the same fabric and takes only a part of it
    Observability
    throughput per traffic type at the Prometheus endpoint

In a cluster, availability is not a property of a node, but a property of the group. What it can withstand is defined by the scaling tier — and what it can still withstand during a maintenance window follows from that by subtraction.

What happens after the loss of a node

  1. 1

    Detect failure

    under 30 seconds

    The remaining nodes lose the heartbeat of the failed one and form a new quorum. Until that is established, nothing is switched over — a briefly disrupted network initially looks like a dead node, and a premature restart of duplicate running guests would be the more expensive mistake.

  2. 2

    Restart guests

    1 to 5 minutes per guest

    The guests of the failed node restart on the remaining ones. Their data is fully available — it resides in the redundancy of the pool, not on the lost device. The restart is a cold start, not a migration: whatever was in the memory is lost.

  3. 3

    Read remotely instead of locally

    Until data locality is restored, the newly started guests read over the network. The cluster is fully usable during this time, but slower — and this is the most visible part of a node failure, not the restart itself.

  4. 4

    Restore redundancy

    The missing blocks are reconstructed from the remaining ones and distributed across the other nodes. Only then can the cluster fully tolerate the next failure again. The run does not need a replacement node — it needs free space on the existing ones.

  5. 5

    Add replacement node

    in the next maintenance window

    The replaced node joins the pool and takes over its share. This is no longer an emergency step, but a redistribution during operations — the cluster is already fully redundant again at this point.

How long recovery takes

Three variables determine the duration: what the failed node carried, how many existing blocks are read for each lost one (mirroring 1+1: 1 block, Erasure coding 8+2: 8 blocks), and how many nodes share the work. A run takes about 28% of the storage fabric, the remaining 72% are left for operations; calculated back from the single value the catalogue states: 6 hours for a PC-L128 with 6 nodes and full allocation.

  • Two-way mirroring (RF2)Stripe 1+1
    Nodes
    3–5
    Read amplification
    1 : 1
    Network load
    7.68 TB
    Duration with fewest nodes
    3 h
    Duration with most nodes
    1.5 h
  • Erasure coding 4+2Stripe 4+2
    Nodes
    6–11
    Read amplification
    4 : 1
    Network load
    19.2 TB
    Duration with fewest nodes
    3 h
    Duration with most nodes
    1.5 h
  • Erasure coding 6+2Stripe 6+2
    Nodes
    12–17
    Read amplification
    6 : 1
    Network load
    26.88 TB
    Duration with fewest nodes
    1.9 h
    Duration with most nodes
    1.3 h
  • Erasure coding 8+2Stripe 8+2
    Nodes
    18–24
    Read amplification
    8 : 1
    Network load
    34.56 TB
    Duration with fewest nodes
    1.6 h
    Duration with most nodes
    1.2 h

The network load is reading and writing combined — which is why it shows a multiple of the 3.84 TB that the node itself carried. Wider stripes do not prolong the run: they distribute it across more nodes, and the two largely cancel each other out. What really prolongs it is a more densely populated class — the worst case in the catalogue is 3.2 h.

A rolling maintenance window, step by step

The window is in the catalogue: Tue 02:00–06:00 CET, rolling with data redistribution before each restart. The difference to a cluster with external storage lies in the third step — before a node restarts, the pool must be rebalanced.

  1. Announcement

    10 calendar days in advance

    The window, scope and sequence of the nodes are fixed in advance. A window without changes to your cluster will be cancelled, not silently run through empty.

  2. Evacuate guests

    5 to 25 minutes per node

    The running VMs migrate to the remaining nodes. Up to this point, the process does not differ from a cluster with an external datastore.

  3. Redistribute data

    20 minutes to several hours per node

    This step only exists here: before a node restarts, the pool is rearranged so that no redundancy level depends solely on it. The duration depends on the node's usage, not on the number of guests.

  4. Firmware and cluster software

    20 to 40 minutes per node

    BIOS, storage controllers, network cards and the cluster software are updated to the target version. After that, the node restarts and rejoins the pool.

  5. Catch up, then the next one

    rolling

    The returned node catches up on what was written during its absence. Only when the pool is fully redundant again does the next node begin. That is why the window length grows with the cluster size — and why the cluster remains usable throughout.

Six identical nodes in an open rack; the third from the top is pulled far out on rails, the status lights of all six are green.
A node in maintenance or failed: the pool sustains it as long as the tier's tolerance allows — in the minimum cluster a, from 6 nodes two. ENTRONYX CLOUD handles the replacement during the maintenance window, while the cluster remains usable.

Recovery per node class

Loss of a node in the smallest erasure coding cluster — the worst case of this tier. The duration increases with the population, not with the price: PC-X256 carries 3.84 TB per node.

PC-M64
3 h · 10 Gbit/s
PC-M64i
6 h · 10 Gbit/s
PC-L128
6 h · 10 Gbit/s
PC-X256
1.2 h · 25 Gbit/s

The storage fabric is listed next to it because it significantly determines the duration: two classes with the same population but different fabric take different amounts of time.

Tolerance during maintenance

3 to 5 nodes
1 outage · 0 in maintenance
from 6 nodes
2 outages · 1 in maintenance
Free nodes per stripe
1 to 8
Maximum size
24 nodes

The right number in the first two rows is the remaining failure tolerance while a node is scheduled out of operation — the tolerance of the tier minus one.

Expansion across a tier boundary

An additional node initially only brings cores, memory and raw capacity. The usable portion only changes at a tier boundary — but then for the entire inventory, which the pool transcodes in the background: throttled, alongside operations, and the new usable capacity is only available afterwards. If you are planning the step from 5 to 6 nodes, you are therefore planning not only the delivery, but also the transcoding.

Upper limit of stripe width. Further widening would increase the recovery load beyond an acceptable level.

No external primary storage, but a way out

The fast datastore classes cannot be booked under hyperconvergence — not by mistake, but because the primary storage is already in the cluster. What is added externally is capacity for cold data and replication to a second region.

Datastore Capacity

iSCSINFS 4.1

NVMe cache in front of HDD capacity — for file services and quiet volumes.

Price per GB and month
€0.0290$0.0336
Snapshots per GB per month
€0.0060$0.0070
Minimum and maximum quantity
8 TB – 512 TB
Redundancy
Erasure coding 8+2 in the location, cache triply mirrored
Backup window
Daily 21:00–06:00 CET, crash-consistent

Restriction: The NVMe cache covers 8% of the datastore size. Working sets beyond this pass through to the HDD tier; the 4 ms latency guarantee then no longer applies.

Datastore Archive

NFS 4.1

Cold capacity for powered-off guests, templates and exports.

Price per GB and month
€0.0068$0.0079
Snapshots per GB per month
€0.0009$0.0010
Minimum and maximum quantity
20 TB – 2048 TB
Redundancy
Erasure coding 8+2, second copy in another EU region
Backup window
Weekly Sat 20:00–08:00 CET

Restriction: Not approved as an operational datastore. Running VMs must not reside here — the class accommodates powered-off guests, templates and OVA exports.

Recovery to a second region

The pool's protection method covers node failure, not location failure. For this, 3 tiers are available; the recovery page shows which of these has a target at the selected location. The example total is calculated with 5.41 TB of protected capacity — the usable capacity of the minimum cluster.

  • Secured recoverySocket €149.00$172.84
    RPO · maximum data loss
    24 h
    RTO · maximum recovery time
    8 h
    Replication interval
    24 h
    Tests / year
    1
    Target region
    Region in the EU, from 150 km
    per GB and month
    €0.0074$0.0086
    Example total
    €189.06$219.31
  • Pilot LightSocket €490.00$568.40
    RPO · maximum data loss
    4 h
    RTO · maximum recovery time
    4 h
    Replication interval
    4 h
    Tests / year
    2
    Target region
    Region in the same country, from 150 km
    per GB and month
    €0.0190$0.0220
    Example total
    €592.87$687.73
  • Asynchronous replicationSocket €990.00$1,148.40
    RPO · maximum data loss
    15 min
    RTO · maximum recovery time
    2 h
    Replication interval
    15 min
    Tests / year
    4
    Target region
    Region in the same country, from 150 km
    per GB and month
    €0.0340$0.0394
    Example total
    €1,174.08$1,361.93
Secured recovery
The restart uses free capacity in the target region. Without booked capacity reservation, the RTO commitment only applies Mon–Fri 08:00–18:00; outside this time, there is no guaranteed restart time.
Pilot Light
A single host runs in the target region with around 25% of the productive capacity. Full throughput is only available after adding more hosts — lead time up to 90 minutes.
Asynchronous replication
The RPO of 15 minutes applies up to a change rate of 40 MB/s per datastore. Above this, the replication lag grows, and the actual recovery point is further back than committed.

Read more

  • Disaster recovery — replication topology, recovery test and the complete cost model of all tiers.
  • Managed dedicated hosts — where the boundary between platform and guest operations lies and what the tiers above it cost.
  • Security and certificates — tenant separation, certificates and the audit documents for the Private Cloud.

Who patches what

Platform operations are included in the node price and cannot be deselected. They end at a strict boundary: we work below the guest system, you work above it.

ENTRONYX CLOUD patches

  • Hypervisor
  • BIOS and firmware
  • Fabric switches
  • Storage controllers

Firmware maintenance requires a host evacuation. Without a second host in the cluster, this means downtime of up to 45 minutes per maintenance window.

You patch

  • Guest operating system
  • Middleware
  • Application
  • Application data

If you also want to hand over the guest system, book a higher operations level. The boundaries of the levels are on the managed hosts page.

If a node fails unexpectedly, the cluster restores the lost redundancy itself. In the smallest cluster of PC-M64, this takes 3 h. While this is running, the cluster cannot sustain a second failure: the mirroring can handle 1 missing node, not two.

Common questions before setting up a cluster

10 questions that regularly come up before ordering. The answers use the same catalogue values as the sections above.

Can't find your question? The cluster planner shows the pool, usable share and total for any number of nodes.

What is hyperconverged infrastructure (HCI), and what does a cluster cost per month?

Hyperconverged infrastructure puts compute and storage in the same nodes: the local NVMe of all nodes form a shared pool, eliminating the need for a separate datastore. ENTRONYX CLOUD operates the cluster in data centres in Germany, starting from 3 nodes; the monthly price per node includes bare metal and platform operations, the licence is added per core.

Minimum cluster per month, net, storage and private network included€5,109.60$5,927.14

Calculate cluster in the planner

Where are the nodes located?

In data centres in Germany; a cluster is located entirely at one location, a second is only added by the recovery tiers. Not every node class is available at every location — the section on node classes shows which is where.

Node classes and their locations

How quickly does support respond if a host fails?

An outage (P1) receives an initial response in 30 minutes, around the clock, in the Platform operations, which is included in every managed host; from the “Platform and guest operations” level it is 15 minutes. Initial response means: a human has accepted the report, classified it and reported back the processing status — we do not guarantee a resolution time. Reports are made via the ticket in the account; a change without an outage requires the lead time of the operations level, in the Platform operations 48 hours.

Failure and maintenance in the cluster

What is the term, and how do I cancel?

Without a term agreement, the contract runs for one month and renews for another; terms of 12 or 24 months are paid in advance and you receive a 5% or 8% term discount respectively. You can cancel monthly contracts with one day's notice to the end of the term, contracts for 12 or 24 months with 30 calendar days' notice — otherwise they renew for twelve months. After the contract ends, your data remains accessible for 30 calendar days; after that, it is irretrievably deleted.

How much of the storage pool is usable?

Between 47% and 76% of the raw capacity, depending on the number of nodes: The pool is created from the local NVMe of all nodes, and the protection method follows the cluster size. A cluster of 3 × PC-M64 provides 11.52 TB raw and 5.41 TB usable. With 6 nodes, it is 14.28 TB, because erasure coding then replaces mirroring.

Pool and usable capacity in the cluster planner

How many nodes does a cluster need?

At least 3 and at most 24 nodes. Only from 3 nodes can the pool decide by majority after a network partition which part is allowed to continue writing.

Minimum cluster per month, net, storage and private network included€5,109.60$5,927.14

Minimum cluster, item by item

How is HCI licensed?

Per core and month, with a minimum purchase of 16 cores per socket; on PC-M64, 48 cores per node are billed. The licence covers the entire primary storage, for which a separate datastore is booked under vSphere. Platform operations are included in the node price.

Licence per core and month, net€13.90$16.12

Licence and price per node class

Which protection method applies, and how many node failures can the pool tolerate?

This depends on the number of nodes, not a setting: up to 5 nodes, two-way mirroring (RF2) applies, from 6 nodes Erasure Coding, starting with 4+2 and up to 8+2 from 18 nodes. Mirroring tolerates one node failure, Erasure Coding two. If a tier limit is exceeded, the pool recodes the existing inventory in the background.

Protection method and usable share per tier

What happens if a node fails?

Its guests restart on the remaining nodes, as a cold start in 1 to 5 minutes per guest; their data lies completely in the redundancy of the pool. Afterwards, the cluster rewrites the missing redundancy itself; in the smallest cluster of PC-M64, this takes 3 h. Until then, the minimum cluster cannot tolerate a second failure.

What happens after the loss of a node

What is the difference between HCI and vSphere with an external datastore?

The location of the primary storage: under HCI, the local NVMe of all nodes form a shared pool, under vSphere a datastore is booked separately. The rest follows from this — HCI starts at 3 instead of 2 nodes, and capacity only grows via additional nodes, whereas a datastore grows individually. In the minimum cluster, HCI is more expensive than the same cluster under vSphere with an equally large Datastore NVMe Performance.

Additional monthly amount, net€24.87$28.85

HCI cluster versus cluster with external storage

Tailor cluster, check capacity

Node class, quantity, supplementary storage, recovery level and operations level are side by side in the configurator. The total is calculated using the same rates as this page — and the usable capacity is tracked, because it changes with every number of nodes.

Per node from
€1,690.20$1,960.63
Minimum cluster
€5,109.60$5,927.14
Usable out of the box
5.41 TB
Nodes per cluster
3 – 24