Skip to main content

SCS Flavor Naming Standard

Introduction

This is the standard v3.2 for SCS Release 8. Note that we intend to only extend it (so it's always backwards compatible), but try to avoid changing in incompatible ways. (See at the end for the v1 to v2 transition where we have not met that goal, but at least managed to have a 1:1 relationship between v1 and v2 names.)

Motivation

In OpenStack environments there is a need to define different flavors for instances. The flavors are pre-defined by the operator, the customer can not change these. OpenStack providers thus typically offer a large selection of flavors.

While flavors can be discovered (openstack flavor list), it is helpful for users (DevOps teams), to have a naming scheme that is used across all SCS flavors, so flavor names have the same meaning everywhere.

While not all details will be encoded in the name, the key features should be obvious: Number of vCPUs, RAM, Root Disk. Extra features are important as well: There will be flavors with GPU support, fast disks for databases, memory-heavy applications, and other useful aspects of an instance.

It may also be important to make the CPU generation clearly recognizable, as this is always a topic in discussions with customers.

Note that not all relevant properties of flavors can be discovered; creating a specification to address this is a separate but related effort to the name standardization. Commonly used infrastructure-as-code tools do not provide a way to use discoverability features to express something like "I want a flavor with 2 vCPUs, 8GiB of RAM, a local 20GB SSD disk and Infiniband support, but I don't care whether it's AMD or intel" in a reasonable manner. Using flavor names to express this will thus continue to be useful, and we don't expect the need for standardization of flavor names to go away until the commonly used IaC tools work on a higher abstraction layer than they currently do.

Design Considerations

Type of information included

From discussions of our operators with their customers we learned that the following characteristics are important in a flavor description:

TypeDescription
GenerationCPU Generation
Number of CPUNumber of vCPUs - suffixed by L,V,T,C (see below)
Amount of RAMAmount of memory available for the VM
Performance ClassAbility to label high-performance CPUs, disks, network
CPU TypeX86-intel, X86-amd, ARM, RISC-V, Generic
"bms"Bare Metal System (no virtualization/hypervisor)

This list is likely not comprehensive and will grow over time.

Rather than using random names s5a.medium and assigning a discrete set of properties to them, we wanted to come up with a scheme that allows to systematically derive names from properties and vice versa. The scheme allows for short names (by not encoding all details) as well as very detailed longer names.

Complete Proposal for systematic flavor naming

PrefixCPUs & SuffixRAM[GiB]optional: Disk[GB]&typeopt: extensions
SCS-NL/V/T/C[i]-N[u][o][-[Mx]N[n/h/s/p]][_EXT]

Note that N and M are placeholders for numbers here. The optional fields are denoted in brackets (and have opt: in the header). See below for extensions.

Note that all letters are case-sensitive.

Typical flavor names look like SCS-4V-16-50 for a flavor with 4vCPUs (with limited oversubscription), 16GiB RAM and a 50GB disk (of unspecified type).

Proposal Details

[REQUIRED] CPU Suffixes

Next to the number of vCPUs, these vCPUs need to be characterized to describe their nature.

SuffixMeaning
Cdedicated Core
Tdedicated Thread (SMT)
VvCPU (oversubscribed)
LvCPU (heavily oversubscribed)

Baseline

Note that vCPU oversubscription for a V vCPU should be implemented such, that we can guarantee at least 20% of a core in >99% of the time; this can be achieved by limiting vCPU oversubscription to 5x per core (or 3x per thread when SMT/HT is enabled) or by more advanced workload management logic. Otherwise L (low performance) instead of V must be used. The >99% is measured over a month (1% is 7.2h/month).

Note that CPUs should use latest microcode to protect against CPU vulnerabilities (Spectre, Meltdown, L1TF, etc.). In particular,

  • microcode must be updated within less than a month of a new release; for CVSS scores above 8, providers should do it in less than a week.
  • all mitigations that are enabled by default in the Linux kernel and the KVM hypervisor should be enabled,
  • CPUs that are susceptible to L1TF (intel x86 pre Cascade Lake) should have hyperthreading disabled OR (in the future) use core scheduling implementations that are deemed to be secure by the SCS security team.

That is to say, except when the suffix i is used, the provider commits itself to implementing the appropriate mitigations if and when they become available, within the timeframes mentioned above.

If a provider does not want to commit to deploying available microcode fixes and upstream kernel/hypervisor updates within a month or if the provider wants to enable hyperthreading on compute hosts despite having CPUs susceptible to L1TF there (and no SCS-accepted core-scheduling mechanism is used for mitigation), the flavors must be declared insecure with the i suffix (see below).

Higher oversubscription

Must be indicated with a L vCPU type (low performance for > 5x/core or > 3x/thread oversubscription and the lack of workload management that would prevent worst case performance < 20% in more than 7.2h per month).

Insufficient microcode

Not using these mitigations must be indicated by an additional i suffix for insecure (weak protection against CPU vulnerabilities through insufficient microcode, lack of disabled hyperthreading on L1TF susceptible CPUs w/o effective core scheduling or disabled protections on the host/hypervisor).

Examples

  • SCS-2C-4-10n
  • SCS-2T-4-10n
  • SCS-2V-4-10n
  • SCS-2L-4-10n
  • SCS-2Li-4-10n
  • SCS-2-**4-10n - CPU suffix missing
  • SCS-2iT-4-10n - This order is forbidden

[REQUIRED] Memory

Baseline

Cloud providers should use ECC memory. Memory oversubscription should not be used. It is allowed to specify half GiBs (e.g. 3.5), though this should not be done for larger memory sizes (>= 10GiB).

No ECC

If no ECC is used, the u suffix must indicate this.

Enabled Oversubscription

If memory is oversubscribed, you must expose this with the o suffix.

Examples

  • SCS-2C-4-10n
  • SCS-2C-3.5-10n
  • SCS-2C-4u-10n
  • SCS-2C-4o-10n
  • SCS-2C-4uo-10n
  • SCS-2C-4ou-10n - This order is forbidden

[OPTIONAL] Disk sizes and types

Disk sizes (in GB) should use sizes 5, 10, 20, 50, 100, 200, 500, 1000.

Disk typeMeaning
nNetwork shared storage (ceph/cinder)
hLocal disk (HDD: SATA/SAS class)
sLocal SSD disk
pLocal high-perf NVMe

Baseline

Note that disk type might be omitted — the user then can not take any assumptions on what storage is provided for the root disk (that the image gets provisioned to).

It does make sense for n to be requested explicitly to allow for smooth live migration. h typically provides latency advantages vs n (but not necessarily bandwidth and also is more likely to fail), s and p are for applications that need low latency (high IOPS) and bandwidth disk I/O. n storage is expected to survive single-disk and single-node failure.

For specific requirements on the SSD and NVMe disks regarding IOPS and power-loss protection, refer to Decision Record scs-0110-ssd-flavors.

If the disk size is left out, the cloud is expected to allocate a disk (network or local) that is large enough to fit the root file system (min_disk in image). This automatic allocation is indicated with - without a disk size. If the - is left out completely, the user must create a boot volume manually and tell the instance to boot from it or use the block_device_mapping_v2 mechanism explicitly to create the boot volume from an image.

Multi-provisioned Disk

The disk size can be prefixed with Mx prefix, where M is an integer specifying that the disk is provisioned M times. Multiple disks provided this way should be independent storage media, so users can expect some level of parallelism and independence.

Examples

  • SCS-2C-4-10n
  • SCS-2C-4-10s
  • SCS-2C-4-10s_bms_z3
  • SCS-2C-4-3x10s - Cloud creates three 10GB SSDs
  • SCS-2C-4-3x10s_bms_z3
  • SCS-2C-4-10 - Cloud decides disk type
  • SCS-2C-4-10_bms_z3
  • SCS-2C-4-n - Cloud decides disk size (min_disk from image or larger)
  • SCS-2C-4-n_bms_3
  • SCS-2C-4- - Cloud decides disk type and size
  • SCS-2C-4-_bms_z3
  • SCS-2C-4-_bms_z3h_GNa-64_ib
  • SCS-2C-4-_ib
  • SCS-2C-4 - You need to specify a boot volume yourself (boot from volume, or use block_device_mapping_v2)
  • SCS-2C-4_bms_z3
  • SCS-2C-4-3x10 - Cloud decides type and creates three 10GB volumes
  • SCS-2C-4-1.5n - You must not specify disk sizes which are not in full GiBs

Naming policy compliance

Every flavor you offer MUST satisfy the following assertion:

  • If its name starts with SCS-, the name has to conform to the syntax outlined above, and the flavor must at least provide the capabilities indicated by the name.

That is to say:

  • You may offer flavors not following the above scheme, as long as the name does not start with SCS-.

  • You are allowed to understate your performance; for instance, a flavor that satisfies SCS-1C-1.5-8s (1 dedicated core, 1.5 GiB RAM, 8 GiB SSD) may also be named SCS-1T-1-5n (1 dedicated hyperthread, 1 GiB RAM, 5 GiB network volume) or even SCS-1V-1-5. Similarly, you may offer the (v3 mandatory) SCS-2V-4-20s with a SCS-2V-4-20p implementation (using a local NVMe instead of an SSD).

You must be very careful to expose low vCPU guarantees (L instead of V), insecure hyperthreading/microcode i, non-ECC-RAM u, memory oversubscription o. Note that omitting these qualifiers is overstating your security, reliability or performance properties and may be reason for clients to feel betrayed or claim damages. This would prevent SCS compliance and certification; in extreme cases, the SCS project might be forced to work with public statements.

We expect all cloud providers to offer the short, less specific flavor names (such as SCS-8V-32-100). Larger providers that offer more details (using the extension below) are expected to still also offer the short variants for usability and easier portability, even beyond the mandated flavors.

You must not extend the SCS naming scheme with your own extensions; you are encouraged however to suggest extensions that we can discuss and add to the official scheme.

Extensions

Extensions provide a possibility for providers that offer a very differentiated set of flavors to indicate hypervisors, support for hardware/nested virtualization, CPU types and generations, high-frequency models, GPU support and GPU types as well as Infiniband support. (More extensions may be appended in the future.)

Using the systematic naming approach ensures that two providers that offer flavors with the same specific features will use the same name for them, thus simplifying life for their customers when consuming these flavors.

Note that there is no need to indicate all details and extra features this way. Flavors may always perform better or have more features than indicated in a name. Underperformance (CPU suffixes L or i or memory suffixes o and u) on the other hand MUST be indicated in the name; this happens rarely in practice.

For smaller providers, the ability to e.g. differentiate between an AMD Milan and an intel IceLake and exposed the slightly different feature set to customers and have slightly different price points is often not worth the extra effort. This is because having this extra differentiation causes fragmentation of the machines (host aggregates) that can offer these flavors, thus resulting in a lower utilization (as the capacity management will need to have a certain amount of headroom per machine pool to avoid running out of capacity).

Note that it is possible for providers to register both the generic short names and the longer, more detailed names and allow them to use the same set of machines (host aggregates). Note that machines (hypervisors) can be part of more than one host aggregate.

The extensions have the format:

[_hyp][_hwv][_arch[N][h]][_[G/g]X[N][-M][h]][_ib]

Extensions are individually optional, but the ones that are used must appear in the order given in the above line.

Remember that letters are case-sensitive. In case you wonder: Feature indicators are capitalized, modifiers are lower case. (An exception is the uppercase _G for a pass-through GPU vs. lowercase _g for vGPU.)

[OPTIONAL] Hypervisor

Format: _hyp

The default Hypervisor is assumed to be KVM. Clouds that offer different hypervisors or Bare Metal Systems should indicate the Hypervisor according to the following table:

hypMeaning
kvmKVM
xenXen
vmwVMware
hyvHyper-V
bmsBare Metal System

Examples

  • SCS-2C-4-10n
  • SCS-2C-4-10n_bms
  • SCS-2C-4-10n_bms_z3h

[OPTIONAL] Hardware virtualization / Nested virtualization

Format: _hwv

If the instances that are created with this flavor support hardware-accelerated virtualization, this can be reflected with the _hwv flag (after the optional Hypervisor flag). On x86, this means that in the instance, the CPU flag vmx (intel) or svm (AMD) is available. This will be the case on Bare Metal flavors on almost all non-ancient x86 CPUs or if your virtualization hypervisor is configured to support nested virtualization. Flavors without the _hwv flag may or may not support hardware virtualization (as we recommend enabling nesting, but don't require flavor names to reflect all capabilities. Flavors may over-deliver ...)

Examples

  • SCS-2C-4-10 - may or may not support HW virtualization in VMs
  • SCS-2C-4-10_kvm_hwv - kvm with enabled nested virtualization
  • SCS-2C-4-10_hwv - not recommended, but allowed
  • SCS-2C-4-10_bms_hwv - better: bare metal with HW virt support (VMX on intel, SVM on AMD, ...)
  • SCS-2C-4-10_hwv_xen - illegal, wrong ordering

[OPTIONAL] CPU Architecture Details

Format: _arch[N][h]

This extension provides more details on the specific CPU:

  • vendor/architecture (arch)
  • generation (N)
  • frequency (h)

Generation and Vendor

The options for arch are as follows:

Lettervendor/architectureCorresponding image architecture
(none)Generic x86-64x86_64
iIntel x86-64x86_64
zAMD (Zen) x86-64x86_64
aARM v8+aarch64
rRISC-V(not yet listed in Glance)

The generation is vendor specific and can be left out, but it can only be specified in conjunction with a vendor. At present, these values are possible:

Generationi (Intel x86-64)z (AMD x86-64) a (AArch64)r (RISC-V)
0pre Skylakepre Zenpre Cortex A76TBD
1SkylakeZen-1 (Naples)A76/NeoN1 classTBD
2Cascade LakeZen-2 (Rome)A78/x1/NeoV1 classTBD
3Ice LakeZen-3 (Milan)A71x/NeoN2/V2(ARMv9)TBD
4Sapphire RapidsZen-4 (Genoa)AmpereOne (ARMv8.6)TBD
5Sierra Forest(E)Zen-5 (Turin)A72x/NeoN3/V3(Av9.2)TBD
6Granite Rapids(P)TBD

It is recommended to leave out the 0 when specifying the old generation; this will help the parser tool, which assumes 0 for an unspecified value and does leave it out when generating the name for comparison. In other words: 0 has a meaning of "rather old or unspecified".

note

We don't differentiate between Zen-4 (Genoa) and Zen-4c (Bergamo); L3 cache per Siena core is smaller on Bergamo and the frequency lower but the cores are otherwise identical. As we already have a qualifier h that allows to specify higher frequencies (which Genoa thus may use more and Bergamo not), we have enough distinction capabilities. The same applies to Zen-5 (Turin) and Zen-5c (Turin Dense). For intel with the server E-cores (Crestmont), these received their own generation assignment, as the difference to the server P-cores (Redwood Cove) is more significant.

Frequency Suffixes

SuffixMeaning
h>2.75GHz all-core
hh>3.25GHz all-core
hhh>3.75GHz all-core

Examples

  • SCS-2C-4-10n
  • SCS-2C-4-10n_z
  • SCS-2C-4-10n_z3
  • SCS-2C-4-10n_z3h
  • SCS-2C-4-10n_z3hh
  • SCS-2C-4-10n_bms_z
  • SCS-2C-4-10n_bms_z3
  • SCS-2C-4-10n_bms_z3
  • SCS-2C-4-10n_bms_z3h
  • SCS-2C-4-10n_bms_z3hh - Bare Metal, AMD Milan with > 3.25GHz all core freq

[OPTIONAL] GPU support

Format: _[G/g]X[N[-M[h][-V[h]]]]

This extension provides more details on the specific GPU:

  • pass-through (G) vs. virtual GPU (g)
  • vendor (X)
  • generation (N)
  • number (M) of processing units that are exposed (for pass-through) or assigned; see table below for vendor-specific terminology
  • high-frequency indicator (h) for compute units
  • amount of video memory (V) in GiB
  • an indicator for high-bandwidth memory

Note that the vendor letter X is mandatory, generation and processing units are optional.

letter Xvendorprocessing units
NnVidiastreaming multiprocessors (SMs)
AAMDcompute units (CUs)
IIntelexecution units (EUs)

For nVidia, the generation N can be f=Fermi, k=Kepler, m=Maxwell, p=Pascal, v=Volta, t=turing, a=Ampere, l=Ada Lovelace, g=Grace Hopper, ..., for AMD GCN-x=0.x, RDNA1=1, C/RDNA2=2, C/RDNA3=3, C/RDNA3.5=3.5, C/RDNA4=4, ... for Intel Gen9=0.9, Xe(12.1/DG1)=1, Xe(12.2)=2, Arc(12.7/DG2)=3 ... (Note: This may need further work to properly reflect what's out there.)

The optional h suffix to the compute unit count indicates high-frequency GPU compute units. It is not normally recommended to use it except if there are several variants of cards within a generation of GPUs and with similar number of SMs/CUs/EUs. In case there are even more than two variants, the letter h can be duplicated for even higher frquencies.

Please note that there are GPUs from one generation and vendor that have vastly different sizes (or different fractions are being passed to an instance with multi-instance-GPUs). The number M allows to differentiate between them and have an indicator of the compute capability and parallelism. M can not at all be compared between different generations let alone different vendors.

The amount of video memory dedicated to the instance can be indicated by V (in binary Gigabytes). This number needs to be an integer - fractional memory sizes must be rounded down. An optional h can be used to indicate high bandwidth memory (such as HBM2+) with bandwidths well above 1GiB/s.

Example: SCS-16V-64-500s_GNa-14-6h This flavor has a pass-through GPU nVidia Ampere with 14 SMs and 6 GiB of high-bandwidth video memory. Looking through GPU specs you could guess it's 1/4 of an A30.

We have a table with common GPUs in the implementation hints for this standard

[OPTIONAL] Infiniband

Format: _ib

This extension indicates Infiniband networking.

More extensions may be forthcoming and appended in a later revision of this spec.

Extensions need to be specified in the above-mentioned order.

Naming options advice

Note that we expect most clouds to prefer short flavor names, not indicating CPU details or hypervisor types. See above list of standard flavors to get a feeling.

However, more successful providers will often need to differentiate their offerings in response to customer demand and allow customers to request flavors with specific detailed properties. The goal of this proposal is to avoid providers to invent their own names and then refer customers to (currently incompletely standardized) extra_specs or worse a non-machine-readable service descriptions to find out the details.

So a cloud provider might well evolve from offering SCS-8T-16-50 to offering SCS-8T-16-50n, SCS-8T-16-50n_i2 and SCS-8T-16-50n_z2 to specify that he is using network disks and offer a choice b/w intel Cascade-Lake and AMD Rome. We would expect the cloud provider to still offer the generic flavor SCS-8T-16-50 and allow the scheduler (placement service) to pick both more specific types (or just one if e.g. capacity management considerations suggest so). Providers in such cases should ensure that the price of a requested flavor does not depend on the scheduler decisions.

We are looking into the metadefs mechanism and extra_specs to allow customers to ask for specific flavor properties without the need to encode all these flavor details into the flavor name, so the optional pieces may not be needed much. However, there must be a way to request flavor properties without encoding the need into an image — the indirection via an image is considered broken by the SCS team.

Proposal Examples

ExampleDecoding
SCS-2C-4-10n2 dedicated cores (x86-64), 4GiB RAM, 10GB network disk
SCS-8Ti-32-50p_i18 dedicated hyperthreads (insecure), Skylake, 32GiB RAM, 50GB local NVMe
SCS-1L-1u-51 vCPU (heavily oversubscribed), 1GiB Ram (no ECC), 5GB disk (unspecific)
SCS-16T-64-200s_GNa-72-24_ib16 dedicated threads, 64GiB RAM, 200GB local SSD, Infiniband, 72 Passthrough nVidia Ampere SMs
SCS-4C-16-2x200p_a14 dedicated Arm64 cores (A76 class), 16GiB RAM, 2x200GB local NVMe drives
SCS-1V-0.51 vCPU, 0.5GiB RAM, no disk (boot from cinder volume)

Previous standard versions

Previous versions up to version 3.0 contained the list of mandatory/recommended flavors, which has been moved to a standard of its own.

Version 1 of the standard used a slightly different naming syntax while the logic was exactly the same. What is a - in v2 used to be a :; _ used to be -. The reason for the change was certain Kubernetes tools using the flavor names as labels. Labels however are subject to stricter naming rules and in particular don't allow for a :. See PR #190 for a discussion.

Version 1 flavor names can be translated to v2 using the following transformation:

NAMEV2=$(echo "$NAMEV1" | sed -e 's/\-/_/g' -e 's/:/-/g' -e 's/^SCS_/SCS-/')

and the way back can be done with

NAMEV1=$(echo "$NAMEV2" | sed -e 's/\-/:/g' -e 's/_/-/g' -e 's/^SCS:/SCS-/')

For the time being, the validation tools still accept the old names with a warning (despite the unchanged SCS- prefix) unless you pass option -2 to them. They will however not count v1 flavors towards fulfilling the needs against the corresponding v2 mandatory flavor list unless you pass the option -1. In other words: An IaaS infrastructure with the 26 v1 mandatory flavors will produce 26 warnings (for using old flavors) and 26 errors (for missing the 26 mandatory v2 flavors) unless you pass -1 in which case no errors and no warnings will be produced. Registering the 26 mandatory v2 flavor names in addition will result in passing the test with only 26 warnings — unless you specify -2. If you do and want to pass you'll need to remove the old v1 names or rename them to no longer start with SCS-.

Beyond SCS

The Gaia-X provider working group which could have created a superseding standard does no longer exist.

However, we have been reaching out to the OpenStack Public Cloud SIG and the ALASCA members to seek further alignment.

Getting upstream OpenStack support for flavor aliases would provide more flexibility and ease migrations between providers, also providers that don't offer the SCS- flavors.

We also would like to see upstream extra_specs standardizing the discoverability of some properties exposed via the SCS names and work on IaC tooling (terraform ...) to make use of these when selecting a flavor.