Posts Tagged ‘reInvent2025’
[AWSReInvent2025] Maximizing Block Storage Performance for High-Intensity Workloads: A Technical Analysis of io2 Block Express and the Nitro System
Lecturer
Mark Olsen and Jody Berenblatt are distinguished engineering and product leaders at Amazon Web Services, specializing in high-performance block storage. Mark Olsen serves as a Principal Product Manager for Amazon EBS, where he focuses on the architectural evolution of Provisioned IOPS volumes to meet the demands of mission-critical enterprise applications. Jody Berenblatt, a Senior Technical Product Manager, brings extensive expertise in the integration of storage subsystems with the AWS Nitro System and the optimization of storage networking protocols. Their work has been pivotal in the development of io2 Block Express, a storage tier designed to provide SAN-like performance in the cloud.
Abstract
This article provides a comprehensive examination of the technical foundations and performance characteristics of high-intensity block storage within the Amazon Elastic Block Store (EBS) ecosystem. Centered on the io2 Block Express architecture, the analysis explores how the integration of the AWS Nitro System, the Scalable Reliable Datagram (SRD) protocol, and Multi-Attach NVMe reservations enables ultra-low latency and high-throughput capabilities for data-intensive workloads such as SAP HANA, Oracle, and Microsoft SQL Server. The discussion details the methodology for managing tail latency, the benefits of decoupled storage architectures, and the operational strategies required to maximize I/O performance in a distributed cloud environment.
Infrastructure Foundations: The Evolution of Provisioned IOPS
The landscape of enterprise computing has shifted toward workloads that demand not only high throughput but also extreme consistency in I/O operations per second (IOPS). For decades, on-premises Storage Area Networks (SANs) were the only viable option for these applications. However, the maturation of Amazon EBS, particularly the transition from io1 to the io2 Block Express architecture, has redefined the capabilities of cloud-native block storage. The fundamental challenge in high-intensity storage is the management of latency, which is often the primary bottleneck for database performance.
In traditional storage models, performance was often tethered to the physical limitations of the disk or the controller. In the modern AWS architecture, the storage is decoupled from the compute instance, connected via a dedicated high-speed network. This separation allows for independent scaling of compute and storage resources but introduces the necessity for highly optimized networking to maintain sub-millisecond latency. The io2 Block Express volumes are engineered to provide up to 256,000 IOPS and 4,000 MB/s of throughput per volume, offering a level of performance that satisfies even the most demanding transactional databases.
Architecture of io2 Block Express: Performance and Durability
The architecture of io2 Block Express represents a paradigm shift in how block storage is provisioned and managed. Unlike standard volumes, io2 Block Express is designed to handle “high-intensity” workloads, defined by their sensitivity to latency and their requirement for high durability. These volumes provide a durability rating of 99.999%, which is a ten-fold improvement over standard io1 volumes. This reliability is achieved through sophisticated replication techniques across multiple physical hardwares within an Availability Zone.
A critical innovation in this architecture is the way it handles I/O operations. By utilizing the Nitro System, the overhead of the hypervisor is removed, allowing the EBS service to communicate directly with the instance’s memory. This “Block Express” layer acts as a high-performance interface that minimizes the processing time required for each I/O request. For applications like SAP HANA, where the speed of logging and data loading is critical, the reduced overhead translates directly into faster business processing cycles.
Networking Innovations: Scalable Reliable Datagram (SRD)
Perhaps the most significant technical advancement in maximizing block storage performance is the implementation of the Scalable Reliable Datagram (SRD) protocol. Traditional TCP protocols, while reliable, are prone to “head-of-line blocking,” where a single lost packet can delay the entire stream of data. In a high-performance storage environment, this creates “tail latency”—spikes in response time that can disrupt database synchronization and performance.
SRD solves this by utilizing multipath routing. Instead of sending data down a single network path, SRD spreads the traffic across as many as 64 different paths simultaneously. If a specific network switch becomes congested or a link fails, the protocol automatically reroutes the data without the latency spikes associated with TCP retransmissions. This protocol is implemented directly in the Nitro Cards, ensuring that the heavy lifting of network management does not consume CPU cycles on the user’s EC2 instance. The result is a more consistent “p99” latency profile, which is essential for maintaining stable performance in clustered environments.
Multi-Attach NVMe Reservations and High Availability
For enterprise applications requiring high availability, the ability for multiple EC2 instances to attach to a single EBS volume is a critical requirement. io2 Block Express supports Multi-Attach, allowing up to 16 Nitro-based instances to access the same volume simultaneously. This feature is particularly valuable for clustered file systems and applications that require shared storage for failover or parallel processing.
To manage concurrent access without data corruption, AWS implemented Multi-Attach NVMe Reservations. Based on the NVMe standard for persistent reservations (similar to SCSI-3 PR), this technology allows one instance to “reserve” the volume, ensuring that only authorized nodes can perform write operations. In the event of an instance failure, the reservation can be quickly cleared and reassigned to a healthy node, minimizing downtime. This mechanism provides the coordination layer necessary for complex deployments like Oracle RAC or SAP environments, where data integrity across multiple nodes is non-negotiable.
Observability and Performance Tuning for Enterprise Workloads
Achieving maximum performance requires a sophisticated approach to observability. Many administrators focus on average latency, but in high-intensity workloads, the “outliers” or tail latency are what truly matter. AWS provides tools such as Amazon CloudWatch and EBS Volume Insights to monitor these metrics in real-time. A key metric is the “Queue Depth,” which represents the number of pending I/O requests for a volume. To reach the full potential of an io2 Block Express volume (e.g., 256,000 IOPS), the application must maintain a sufficient queue depth—often 128 or higher—to keep the storage pipeline full.
// Example AWS CLI command to modify an EBS volume to io2 with high provisioned IOPS
aws ebs modify-volume \
--volume-id vol-0123456789abcdef \
--volume-type io2 \
--iops 100000
Furthermore, the choice of the EC2 instance type is paramount. Performance is not solely a function of the storage volume; the instance must be “EBS-optimized” with sufficient dedicated bandwidth to handle the provisioned throughput. For instance, using an R5b or X2idn instance allows the application to utilize the full 4,000 MB/s throughput offered by Block Express. Failure to match the instance capability with the volume performance will lead to throttling at the instance level, regardless of how many IOPS are provisioned.
Links:
[AWSReInvent2025] Advancements in AWS Infrastructure as Code: A Comprehensive Year-in-Review of CloudFormation and CDK Innovations
Lecturer
The session is delivered by product managers from Amazon Web Services who oversee the development and roadmap of AWS CloudFormation and the AWS Cloud Development Kit.
Abstract
This article provides an exhaustive and detailed retrospective on the notable progress achieved throughout the past year in AWS infrastructure as code services, with particular emphasis on both AWS CloudFormation and the AWS Cloud Development Kit (CDK). It meticulously examines a range of enhancements, including improved validation mechanisms, clearer error diagnostics, expanded construct libraries, seamless integration with artificial intelligence assistance through Model Context Protocol servers, and advanced troubleshooting utilities. The discussion analyzes how these collective innovations substantially elevate deployment reliability, enhance developer productivity, and introduce greater intelligence into infrastructure management practices for organizations of all scales.
The Critical and Enduring Role of Infrastructure as Code in Modern Cloud Architectures
Infrastructure as code has firmly established itself as an indispensable discipline for enterprises striving to achieve consistency, traceability, and accelerated iteration in their cloud operations. AWS CloudFormation offers a robust declarative approach, allowing practitioners to define resources through structured templates in JSON or YAML formats, thereby guaranteeing identical provisioning outcomes across development, staging, and production environments.
Complementing this, the AWS Cloud Development Kit empowers developers with programmatic flexibility, enabling infrastructure definition in familiar programming languages while automatically generating underlying CloudFormation templates. This duality accommodates diverse team preferences and skill sets.
The advancements introduced over the year have strategically bridged these paradigms, delivering unified capabilities that address contemporary challenges related to scale, complexity, and the evolving demands of developer experience in dynamic cloud ecosystems.
Significant Refinements Enhancing AWS CloudFormation Reliability and Practitioner Usability
AWS CloudFormation has benefited from meaningful improvements in change set validation processes, enhanced clarity in error messaging, and more intuitive management of deployment workflows. These refinements work collectively to substantially reduce the frequency of failed deployments by surfacing potential conflicts, resource constraints, or configuration incompatibilities earlier in the provisioning lifecycle.
Furthermore, the introduction of server-side APIs now enables programmatic pre-validation of proposed changes, allowing integration into continuous integration pipelines for automated safeguards that prevent runtime disruptions and promote greater confidence in infrastructure updates.
Substantial Growth and Maturation Within the AWS Cloud Development Kit Ecosystem
The AWS Cloud Development Kit has experienced considerable expansion in supported programming languages and the availability of high-level constructs. Numerous libraries, both community-contributed and AWS-maintained, have progressed from experimental developer preview stages to full general availability, covering an extensive array of common architectural patterns across networking, security, serverless computing, and data processing domains.
This maturation process provides developers with higher-level abstractions that encapsulate established best practices, thereby significantly reducing the amount of boilerplate code required and promoting greater architectural consistency across distributed teams.
Transformative Integration of Artificial Intelligence Assistance Through Model Context Protocol Servers
One of the most pivotal innovations involves the creation of specialized Model Context Protocol servers tailored specifically for CDK and CloudFormation contexts. These servers curate and expose AWS-specific expertise—including recommended practices, construct libraries at various maturity levels, and detailed cloud context information—directly to artificial intelligence-powered coding assistants.
As a result, developers receive highly contextually relevant suggestions that align precisely with AWS service conventions and idioms, dramatically accelerating the creation of secure, efficient, and idiomatic implementations while substantially lowering the cognitive burden associated with recalling intricate service details.
Strengthening Troubleshooting and Validation Tooling for Proactive Issue Resolution
New diagnostic capabilities encompass server-side APIs designed for interrogating deployment states and identifying root causes of issues, complemented by local static analysis utilities that perform early detection of syntax errors within CDK source code.
These tools operate across both programmatic CDK definitions and the generated CloudFormation templates, enabling practitioners to identify and resolve configuration problems well before they manifest during actual deployments.
Community-Driven Construct Libraries and Enhanced Cloud Context Integration
The ecosystem continues to benefit from active contributions spanning AWS internal teams and external community participants, with constructs systematically progressing through alpha evaluation and eventual general availability phases.
Additional cloud context features further enrich artificial intelligence interactions by providing service-specific insights and recommendations.
Practitioners are strongly encouraged to explore dedicated workshops that offer guided paths for understanding and implementing MCP server integration in real-world scenarios.
Measurable Organizational Benefits and Strategic Adoption Considerations
These multifaceted improvements collectively lower entry barriers for effective infrastructure management while delivering tangible advantages. Development teams realize enhanced confidence in deployment outcomes, accelerated onboarding for new members, and improved adherence to evolving architectural standards across projects.
The incorporation of artificial intelligence guidance represents a fundamental paradigm shift toward more intelligent, assisted development experiences that amplify human expertise rather than seeking to replace it.
Looking Toward the Future of Intelligent Infrastructure Orchestration
Continued investment in these areas clearly signals an ongoing commitment to deepening the convergence between programmatic expressiveness and declarative safety, increasingly augmented by artificial intelligence capabilities that guide practitioners toward optimal architectural outcomes.
Organizations that fully leverage these evolving tools position themselves advantageously for sustained operational excellence amid the accelerating complexity of modern cloud environments.
Links:
[AWSReInvent2025] Amazon S3 Performance: Architecture, Design, and Optimization for Data-Intensive Systems
Lecturer
Ian Heritage is a Senior Solutions Architect at Amazon Web Services, specializing in Amazon S3 and large-scale data storage architectures. With deep expertise in performance engineering and distributed systems, Ian Heritage helps organizations design and optimize their storage layers for high-throughput and low-latency applications, including machine learning training and real-time analytics. He is a prominent figure in the AWS storage community, known for his technical deep-dives into S3’s internal mechanics and best practices for performance at scale.
Abstract
This article explores the internal architecture and performance optimization strategies of Amazon S3, the industry-leading object storage service. It provides a detailed analysis of the differences between S3 General Purpose and the newly introduced S3 Express One Zone storage class, highlighting the architectural trade-offs between regional durability and sub-millisecond latency. The discussion covers advanced request management techniques, including prefix partitioning, request routing, and the role of the AWS Common Runtime (CRT) in maximizing throughput. By examining these technical foundations, the article offers practical guidance for architecting storage solutions that can handle millions of requests per second and petabytes of data for modern AI and analytics workloads.
S3 Storage Class Selection for High Performance
The performance of an S3-based application is fundamentally determined by the selection of the storage class. For over a decade, S3 General Purpose (Standard) has been the default choice, offering 99.999999999% (11 9s) of durability by replicating data across at least three Availability Zones. While this provides extreme reliability, the regional replication introduces a baseline latency that may be too high for certain “request-intensive” applications, such as machine learning model checkpoints or high-frequency trading logs.
To address these needs, AWS introduced S3 Express One Zone. This storage class is designed for workloads that require consistent, single-digit millisecond latency. By storing data within a single Availability Zone and utilizing a new, purpose-built architecture, Express One Zone can deliver up to 10x the performance of S3 Standard at a 50% lower request cost. This class is ideal for applications that perform frequent, small I/O operations where the overhead of regional replication would be the primary bottleneck. The choice between Standard and Express One Zone is thus a strategic decision between geographic durability and extreme performance.
Request Routing, Partitioning, and the Scale-Out Architecture
At its core, Amazon S3 is a massively distributed system that scales out to handle virtually unlimited throughput. The key to this scaling is “partitioning.” S3 automatically partitions buckets based on the object keys (names). Each partition can support a specific number of requests: 3,500 PUT/COPY/POST/DELETE requests and 5,500 GET/HEAD requests per second per prefix. For many years, users were advised to use randomized prefixes to ensure even distribution across partitions.
Modern S3 architecture has evolved to handle this automatically, but understanding prefix design remains crucial for performance. When an application’s request rate increases, S3 detects the hot spot and splits the partition to handle the load. However, this process takes time. For workloads that burst from zero to millions of requests instantly, pre-partitioning or using a wide range of prefixes is still a best practice. By spreading data across multiple prefixes (e.g., bucket/prefix1/, bucket/prefix2/), an application can linearly scale its throughput to accommodate massive concurrency, limited only by the client’s network bandwidth and CPU.
Client-Side Optimization with AWS CRT and SDKs
While the S3 service is designed for scale, the performance experienced by the end-user is often limited by the client-side implementation. To bridge this gap, AWS developed the Common Runtime (CRT) library. The CRT is a set of open-source, C-based libraries that implement high-performance networking best practices, such as automatic request retries, congestion control, and most importantly, multipart transfers.
'''
Conceptual example of enabling CRT in the AWS SDK for Python (Boto3)
'''
import boto3
from s3transfer.manager import TransferConfig
'''
The CRT allows for automatic parallelization of large object transfers
'''
config = TransferConfig(use_threads=True, max_concurrency=10)
s3 = boto3.client('s3')
s3.upload_file('large_data.zip', 'my-bucket', 'data.zip', Config=config)
The CRT automatically breaks large objects into smaller parts and uploads or downloads them in parallel. This utilizes the full network capacity of the EC2 instance and mitigates the impact of single-path network congestion. For applications using the AWS CLI or SDKs for Java, Python, and C++, opting into the CRT-based clients can result in a significant throughput increase—often double or triple the speed of standard clients for large files. Additionally, the CRT handles the complexities of DNS load balancing and connection pooling, ensuring that requests are distributed efficiently across the S3 frontend fleet.
Case Study: Optimization for Machine Learning and Analytics
Machine learning training is a premier use case for S3 performance optimization. During the training of large language models (LLMs), hundreds or thousands of GPUs must simultaneously read training data and write model “checkpoints.” These checkpoints are multi-gigabyte files that must be saved quickly to avoid idling expensive compute resources. By combining S3 Express One Zone with the CRT-based client, researchers can achieve the throughput necessary to saturate the high-speed networking of P4 and P5 instances.
In analytics, the use of “Range Gets” is a critical optimization. Instead of downloading an entire 1GB Parquet file to read a few columns, an application can request specific byte ranges. This reduces the amount of data transferred and speeds up query execution. S3 is optimized to handle these range requests efficiently, and when combined with a partitioned data layout (e.g., partitioning by date or region), it enables sub-second query responses over petabytes of data. This architectural synergy between storage class, partitioning, and client-side logic is what allows S3 to serve as the foundation for the world’s largest data lakes.