AWS Launches ARM-Powered Servers & Networking Enhancements
SEATTLE -- Today at AWS re:Invent, Amazon Web Services, Inc. (AWS), an Amazon.com company, announced three new instance offerings for Amazon Elastic Compute Cloud (Amazon EC2):
* A1 instances—powered by custom-designed AWS Graviton processors for scale-out workloads
* P3dn GPU instances—ideal for distributed machine learning and high performance computing applications
* C5n instances—deliver increased network bandwidth for running advanced compute-intensive workloads
The new P3dn GPU and C5n compute optimized instances feature 100 Gbps networking throughput and enable scale-out of distributed workloads like high performance computing (HPC), machine learning training, and data analytics.
AWS also announced the availability of two new networking offerings:
* Elastic Fabric Adapter (EFA)—a network adapter for Amazon EC2 instances that delivers the performance of on-premises HPC clusters with AWS’s elasticity and scalability
* AWS Global Accelerator—a fully managed service that uses AWS's global backbone and edge locations to improve the availability and performance of applications running in one or more AWS regions
“Two of the requests we get most from customers are how can you help us keep lowering our costs for basic workloads, and how can you make it more efficient to run our demanding, scale-out, high performance computing and machine learning workloads in the cloud,” said Matt Garman, Vice President of Compute Services, AWS. “With today’s introduction of A1 instances, we’re providing customers with a cost optimized way to run distributed applications like containerized microservices. A1 instances are powered by our new custom-designed AWS Graviton processors with the Arm instruction set that leverages our expertise in building hyperscale cloud platforms for over a decade. For scale-out distributed workloads, our new P3dn instances and C5n instances offer 100 Gbps networking performance to speed distributed machine learning training and high performance computing. These new instance launches expand what's already the industry’s most powerful and cost-effective computing platform to meet the needs of new and emerging workloads.”
New A1 instances feature custom-designed AWS Graviton processors that deliver significant cost saving for scale-out workloads
Customers increasingly run a diverse set of workloads in the cloud and are looking for solutions that lower costs without compromising performance. Although general purpose processors continue to provide great value for many workloads, new and emerging scale-out workloads like containerized microservices and web tier applications that do not rely on the x86 instruction set can gain additional cost and performance benefits from running on smaller and modern 64-bit Arm processors that work together to share an application’s computational load. Available today, A1 instances feature a custom-designed processor (Graviton) that leverages AWS’s extensive expertise in systems design and cloud infrastructure, making Arm processors available in the cloud for the first time. With A1 instances, customers will benefit from up to a 45 percent cost reduction (compared to other Amazon EC2 general purpose instances) for scale-out workloads. A1 instances are supported by several Linux distributions, including Amazon Linux 2, Red Hat, and Ubuntu, as well as container services, including Amazon Elastic Container Service (Amazon ECS) and Amazon Elastic Container Service for Kubernetes (EKS).
SmugMug is a paid image sharing, image hosting and online video platform. “We are constantly striving to make our service more affordable for our fast growing customer base,” said Don MacAskill, CEO and Chief Geek of SmugMug. “We run all our workloads in the cloud and controlling costs is important. Our technology stack consists mainly of PHP and is easily portable to the Arm architecture. The new A1 instances are a great fit for our web tier and provide a 40 percent cost reduction benefit.”
New P3dn and C5n Instances feature100 Gbps networking performance, accelerating computationally demanding workloads
Today, many customers are turning to AWS compute-optimized C5 instances and GPU-powered P3 instances for some of the most compute intensive workloads in the cloud. From machine learning training, to HPC applications like computational fluid dynamics and weather simulations, to video encoding, these workloads benefit from powerful processors and high speed networking. AWS is the first cloud provider to deliver 100 Gbps networking performance in a secure, scalable, and elastic manner so that customers can use it not just for HPC, but also for analytics, machine learning, big data, and data lake workloads with standard drivers and protocols.
* P3dn instances (available next week) will be the most powerful GPU instances in the cloud for machine learning training. P3 instances already help customers accelerate machine learning model training time from several days to a few hours, and with the 100 Gbps networking performance of the new larger size P3dn instances, customers can further lower their training times to less than an hour by distributing their machine learning workload across multiple GPU instances. Since launching P3 instances in October 2017, customer demand for higher performance compute has increased as machine learning adoption grows and customers apply it to tackle more complex applications. The new P3dn instances deliver a 4X increase in network throughput compared to existing P3 instances, providing up to 100 Gbps of networking throughput, fast NVMe instance storage, custom Intel CPUs with 96 vCPUs and support for AVX512 instructions, and NVIDIA Tesla V100 GPUs each with 32 GB of memory. This enables developers to linearly scale their model training performance across multiple instances, accelerate preprocessing and remove data transfer bottlenecks, and rapidly improve the quality of their machine learning models.
fast.ai is an organization dedicated to making the power of deep learning accessible to all. “We have been using Amazon EC2 P3 instances to demonstrate how machine learning models can be trained quickly, easily, and inexpensively. In August 2018, we were able to train Imagenet to 93% accuracy in just 18 minutes, using 16 P3 instances for only around $40,” said Jeremy Howard, Founding Researcher, fast.ai. “We look forward to using the new P3dn instances with their 100 Gbps of network throughput and increased compute power for scaling out to include more instances and further decrease the time it takes to train our models.”
* C5n instances (available today) significantly increase the maximum throughput performance available in AWS’s compute-intensive instance family. C5 instances offer up to 25 Gbps of network bandwidth addressing the requirements of a wide range of workloads, but highly distributed and HPC applications can benefit from even higher network performance. C5n instances offer 100 Gbps of network bandwidth, providing four times as much throughput as C5 instances. This performance increase delivered with C5n instances enables previously network bound applications to scale up or scale out effectively on AWS. Customers can also take advantage of higher network performance to accelerate data transfer to and from Amazon Simple Storage Service (Amazon S3), reducing the data ingestion wait time for applications and speeding up delivery of results.
Teradata transforms how businesses work and people live through the power of data. "Teradata IntelliCloud, our as-a-service offering for analytics at scale, enables customers to seamlessly scale up and out to meet their workload requirements," said Abhishek Lal, Director of Strategic Offering Management for Cloud at Teradata. "Teradata software demands extremely high I/O to maximize its potential. With a 4X improvement in network performance, we expect Amazon EC2 C5n instances to significantly improve throughput for IntelliCloud, empowering customers to generate analytic insights and business-changing answers at ever-faster rates."
Low-latency Elastic Fabric Adapter aids migration of HPC workloads to AWS
Customers typically rely on fixed-size, on-premises HPC systems. Because HPC systems are capital-intensive and expensive, companies tend to under-procure this capacity, resulting in long wait times that decrease productivity, limit experimentation, and delay critical project work. Moreover, customers are held captive to the available hardware capabilities and technologies until the next infrastructure refresh cycle – forcing developers to adapt their applications to the infrastructure instead of the other way around. EFA (available in preview today) enhances the performance of inter-instance communications that is critical for scaling HPC applications, providing customers the performance they expect from on-premises HPC clusters in the cloud, with the added benefits of AWS’s elasticity and scalability. EFA is integrated with the Message Passing Interface (MPI), which allows HPC applications to scale to tens of thousands of CPU cores without any modification. EFA is available on Amazon EC2 P3dn and C5n instances, and it will be enabled on additional Amazon EC2 instance types in 2019, giving customers added flexibility to choose the right compute configuration for their workloads on-demand, without any upfront planning.
AWS Global Accelerator improves availability and performance of globally distributed applications
As customers scale for a larger and more geographically-diverse set of users, they have to operate with better availability and improved performance. These customers deploy applications in multiple AWS regions for better performance, but this means they have to route users to the right healthy application endpoint. Further, they must regularly scale up or down the application’s endpoints, each with their own IP address, in response to application failures, performance testing, or spikes in traffic. As their applications scale, they also have to update every client that connects to the application (typically done via a Domain Name Service), which increases the time it takes for these clients to discover the new endpoints. AWS Global Accelerator (available today) improves the availability and performance of applications and makes it simple to direct internet traffic from users to application endpoints running in multiple AWS regions. It uses AWS’s vast, highly available and congestion-free global network backbone and edge locations to direct internet traffic from users to the application endpoints. Clients are directed to the right application endpoint based on their geographic location, application health, and customer-configurable routing policies. AWS Global Accelerator also allocates a set of static Anycast IP addresses that are unique per application and do not change, thus removing the need to update clients as the application scales. Application endpoints are continuously monitored and AWS Global Accelerator only directs clients to healthy endpoints without any need to change client configuration.
Located in Munich, Germany, Team Internet AG is a leading provider of services in the direct navigation search market. “Today we host our applications in five AWS Regions and plan to continue to expand to new regions to accelerate performance and achieve high availability for our users,” said Markus Ostertag, CEO, Team Internet AG. “With AWS Global Accelerator’s anycast of IP addresses and the intelligent edge network, we can now respond to application failures or configuration updates in a matter of seconds, to ensure maximum uptime and lowest latency for our users. We are really excited about AWS Global Accelerator’s support for most network layer protocols, since we have UDP-based workloads.”