Elastic Compute Service:Instance families

Enterprise-level vs. shared instances

• Enterprise-level instance families suit production workloads of all sizes, including databases, video encoding, and data analytics.

• Shared instance families (entry-level) target small and medium-sized websites or individual developers. Shared instance types do not provide stable computing performance but cost less because they share underlying resources.

Arm-based use cases

Arm-based instance families suit containers, microservices, website and application servers, high-performance computing, and CPU-based machine learning.

Instance size reference

The size label maps to vCPU count as follows:

Worked examples

The following examples show how to decode common instance type names.

Series reference for x86 and Arm instance families

The series letter indicates the workload category and defines the vCPU-to-memory ratio:

Suffix reference

Suffixes fall into three categories: processor type, feature enhancement, and role specification.

Processor suffixes

Feature suffixes

Role suffixes

Series reference for heterogeneous, Bare Metal, and SCC instance families

These instance families use additional series prefixes:

Heterogeneous computing series

Bare Metal and SCC series

GPU generation and model suffixes

For GPU-accelerated instance families, the digit indicates the GPU architecture, and the suffix letter identifies the GPU model:

The digit in the suffix (6 or 7) indicates the GPU architecture. The suffix letter identifies the specific GPU model or variant within that architecture generation.

Processor

The physical CPU model powering the instance.

• Base frequency -- the standard clock speed at which the processor runs under typical workloads, without overclocking or special optimization.

• Turbo frequency -- the maximum clock speed the processor can temporarily achieve when demand increases.

vCPU

Each vCPU of x86-based instance types corresponds to a hyperthread of a processor core. Each vCPU of Arm-based instance types corresponds to a physical processor core. Enterprise-level instance types in both architectures deliver stable performance using dedicated resources.

Burst performance (burstable instances only)

• Average baseline CPU performance -- the CPU capacity continuously available to the instance.

• CPU credits per hour -- the rate at which the instance earns CPU credits. The credits-per-hour value in instance type tables reflects the total credits earned by all vCPUs on the instance.

• Maximum CPU credit balance -- the maximum credits the instance can accumulate within a 24-hour period. CPU credits are valid for up to 24 hours.

Memory

• Memory -- stores data for rapid access while the instance runs. Data in memory is lost when the instance shuts down or restarts.

• Persistent memory -- can serve as memory or local storage. Listed separately in instance type tables for families that support it.

• Encrypted memory -- provided through Intel Software Guard Extensions (SGX) technology.

Network bandwidth

• Baseline bandwidth -- the maximum amount of data that can be transmitted over a network connection in a specific period of time. This is the initial bandwidth setting for the network connection.

• Burst bandwidth -- the peak bandwidth the instance can temporarily reach by consuming network burst credits. Available on specific instance types in 6th-generation or later families. No SLA commitment.

• Full-duplex bandwidth -- supported on 7th-generation or later families. Inbound and outbound bandwidth each reach their specified values independently, without affecting each other.

All network specifications are measured in a forwarding-only test environment. Actual performance varies based on workload type, packet size, connection type (persistent or short-lived), image version, and networking model. Run stress tests on your actual workload to validate instance type selection.

Packet forwarding rate (PPS)

The sum of maximum inbound and outbound packet forwarding rates, measured in packets per second (PPS).

Measured in a forwarding-only test environment. Actual performance may vary.

Connections

A connection (session) is defined by a 5-tuple: source IP address, destination IP address, source port, destination port, and transport protocol. ECS connection metrics include TCP, UDP, and Internet Control Message Protocol (ICMP) connections.

NIC queues

The maximum number of network interface controller (NIC) queues supported per Elastic Network Interface (ENI). More NIC queues enable more efficient distribution of network traffic, reducing packet wait times, packet loss, and latency.

Elastic Network Interfaces (ENIs)

The maximum number of ENIs that can be bound to an instance. Bind secondary ENIs to different instances for flexible, scalable network configurations.

Elastic RDMA Interfaces (ERIs)

The maximum number of Elastic RDMA Interfaces (ERIs) that can be bound to an instance. ERIs are RDMA-capable ENIs that reuse the networks to which ENIs belong.

Jumbo frames support

Indicates whether the instance type supports Jumbo Frames. Alibaba Cloud supports jumbo frames with an 8,500-byte Ethernet payload.

Private IPv4 addresses per ENI

The maximum number of private IPv4 addresses supported per ENI.

IPv6 addresses per ENI

The maximum number of IPv6 addresses supported per ENI.

I/O optimization

I/O optimization provides improved network and storage performance for instances and cloud disks.

Local storage

Local storage (local disks) refers to disks attached to the physical machine hosting the ECS instance. Local storage provides temporary block storage, cannot be separately created, and is measured in GiB.

Disk bandwidth

• Baseline disk bandwidth -- the sustained disk bandwidth guaranteed by the SLA for cloud disks.

• Burst disk bandwidth -- the peak disk bandwidth the instance can temporarily reach. No SLA commitment.

Disk input/output operations per second (IOPS)

• Baseline IOPS -- the sustained IOPS guaranteed by the SLA for cloud disks.

• Burst IOPS -- the peak IOPS the instance can temporarily reach. No SLA commitment.

vTPM

Virtual Trusted Platform Module (vTPM) enables trusted computing capabilities. Trusted Platform Modules (TPMs) or Trusted Cryptography Modules (TCMs) on the underlying physical servers serve as trusted computing bases (TCBs), providing tamper-protected, trusted boot for instances. vTPM can be used to measure the critical components of the boot chain of instances.