SCC instance families

更新时间:
复制 MD 格式

This topic describes the features of Super Computing Cluster (SCC) instance families of Elastic Compute Service (ECS) and lists the instance types of each instance family.

Note

Introduction

SCC is based on ECS Bare Metal Instance and uses high-speed Remote Direct Memory Access (RDMA) based interconnections to significantly improve the network performance and acceleration ratio of large-scale clusters. SCC has all the benefits of ECS Bare Metal Instance and provides high-bandwidth, low-latency networks.

SCC is used in scenarios such as high-performance computing (HPC), AI, machine learning, scientific and engineering computing, data analytics, and audio and video processing. In a cluster, nodes are connected through high-throughput, low-latency RDMA networks to serve high-concurrency workloads of applications, such as HPC and AI applications. RDMA over Converged Ethernet (RoCE) networks rival InfiniBand networks in terms of connection speed and can support more Ethernet-based applications.

SCC can work with other Alibaba Cloud computing services, such as ECS and Elastic GPU Service, to provide ultra-high-performance parallel computing resources for Alibaba CloudE-HPC, making cloud-based super computing possible.

HPC optimized instance families contain the following:

Comparison of SCCs, physical machines, and virtual machines

The following table compares the features of SCCs, physical machines, and virtual machines. In the table, Y indicates that the feature is supported, N indicates that the feature is not supported, and N/A indicates that the feature is not applicable.

Feature type

Feature

SCC

Physical machine

Virtual machine

Automated O&M

Minute-level delivery

Y

N

Y

Compute

Zero performance loss

Y

Y

N

Zero feature loss

Y

Y

N

Zero resource contention

Y

Y

N

Storage

Compatibility with ECS disks

Y

N

Y

Boot from system disks

Y

N

Y

Quick reset of system disks

Y

N

Y

Use of ECS images

Y

N

Y

Cold migration between physical and virtual machines

Y

N

Y

No need to install operating systems

Y

N

Y

No need for local RAIDs and better data protection for disks

Y

N

Y

Network

Compatibility with virtual private clouds (VPCs)

Y

N

Y

No communication bottlenecks between physical and virtual machine clusters in VPCs

Y

N

Y

Management

Compatibility with existing ECS management systems

Y

N

Y

Consistent user experience on features such as Virtual Network Computing (VNC) with that on virtual machines

Y

N

Y

Out-of-band (OOB) network security

Y

N

N/A

sccg7, general-purpose SCC instance family

  • Introduction: This instance family provides all features of ECS Bare Metal Instance. For more information, see Overview of ECS Bare Metal Instance families.

  • Supported scenarios:

    • Large-scale machine learning training

    • Large-scale high-performance scientific computing and simulations

    • Large-scale data analytics, batch processing, and video encoding

  • Compute:

    • Offers a CPU-to-memory ratio of 1:4.

    • Uses 2.9 GHz Intel® Xeon® Platinum 8369 (Ice Lake) processors that deliver an all-core turbo frequency of 3.5 GHz.

  • Storage:

  • Network:

    • These instances support IPv4 and IPv6. For information about IPv6 communication, see IPv6 communication.

    • Supports both RoCE networks and VPCs. RoCE networks are dedicated to RDMA communication.

sccg7 instance types

Instance type

vCPUs

Physical cores

Memory (GiB)

Network baseline bandwidth (Gbit/s)

Packet forwarding rate (pps)

RoCE network bandwidth (Gbit/s)

ENIs

ecs.sccg7.32xlarge

128

64

512.0

100

24,000,000

200

32

sccc7, compute-optimized SCC instance family

  • Introduction: This instance family provides all features of ECS Bare Metal Instance. For more information, see Overview of ECS Bare Metal Instance families.

  • Supported scenarios:

    • Large-scale machine learning training

    • Large-scale high-performance scientific computing and simulations

    • Large-scale data analytics, batch processing, and video encoding

  • Compute:

    • Offers a CPU-to-memory ratio of 1:2.

    • Uses 2.9 GHz Intel® Xeon® Platinum 8369 (Ice Lake) processors that deliver an all-core turbo frequency of 3.5 GHz.

  • Storage:

  • Network:

    • These instances support IPv4 and IPv6. For information about IPv6 communication, see IPv6 communication.

    • Supports both RoCE networks and VPCs. RoCE networks are dedicated to RDMA communication.

sccc7 instance types

Instance type

vCPUs

Physical cores

Memory (GiB)

Network baseline bandwidth (Gbit/s)

Packet forwarding rate (pps)

RoCE network bandwidth (Gbit/s)

ENIs

ecs.sccc7.32xlarge

128

64

256.0

100

24,000,000

200

32

sccgn7ex, GPU-accelerated compute-optimized SCC instance family

  • Introduction: The sccgn7ex family provides high-bandwidth SCC instances. Alibaba Cloud developed this family to meet the growing demand for large-scale AI training. Multiple bare metal servers are interconnected through a third-generation RDMA SCC network. This network supports an interconnection bandwidth of 800 Gbit/s. You can scale the number of clusters based on your training needs to quickly meet the demands of large-scale AI parameter training.

  • Scenarios: Ultra-large-scale AI training.

  • Compute:

    • These instances support NVSwitch and provide up to 312 TFLOPS of computing power (TF32).

    • The processor-to-memory ratio is 1:8.

    • Processor: Third-generation Intel® Xeon® 8369 scalable processors (Ice Lake). These processors have a base frequency of 2.9 GHz, an all-core turbo frequency of 3.5 GHz, and support the PCIe 4.0 interface.

  • Storage:

  • Network:

    • These instances support IPv4 and IPv6. For information about IPv6 communication, see IPv6 communication.

    • Supports only VPC.

    • High network performance with a packet forwarding rate of 24 million PPS.

    • sccgn7ex instances support an interconnection bandwidth of 800 Gbit/s (4 × dual-port 100 Gbit/s RDMA). The instances support GPUDirect. Each GPU is directly connected to a 100 Gbit/s network port.

The following table lists the instance types and specifications for the sccgn7ex family.

Instance type

vCPU

Memory (GiB)

GPU memory (GB)

Base network bandwidth (Gbit/s)

Packet forwarding rate (PPS)

RoCE network (Gbit/s)

Elastic Network Interfaces (ENIs)

ecs.sccgn7ex.32xlarge

128

1024

80 GB × 8

64

24 million

800

15

Billing methods

SCCs support the pay-as-you-go and subscription billing methods. For more information, see Overview of billing methods.