Azure IaaS: Deploy high-performance workloads with a system-level approach

Cloud performance isn’t just about individual resources anymore; it’s all about how compute, storage, and networking interact. Azure IaaS adopts a system-level approach that helps businesses achieve reliable, scalable performance for AI, cloud-native, and critical workloads.

This blog post is part of the Azure IaaS series, where we share best practices and advice to help you establish a trusted infrastructure platform, focusing on performance, resilience, security, scalability, and cost-effectiveness.

Performance is a key factor in the success or failure of applications in the cloud. Whether you’re training AI models, scaling a Kubernetes setup, or managing a crucial database, performance is about more than just deciding on CPU, storage, or networking. It’s about how all three components work together, and it demands a comprehensive strategy.

A lot of organisations still tackle performance by boosting resources—like opting for larger VMs, faster disks, or greater network bandwidth. However, modern workloads can behave unpredictably, making this approach less effective. Bottlenecks can shift quickly, with a database potentially limited by storage latency at one moment and network bandwidth just moments later. An AI pipeline might stall not because of compute restrictions, but purely because data can’t move swiftly enough between nodes.

This shift is crucial; performance in the cloud is no longer a resource-centric issue, but a system-wide challenge. Azure recognises this by engineering performance directly into its platform, allowing clients to achieve steady, scalable results without the need to manually adjust every single component.

Rethinking cloud performance

Performance is no longer solely about achieving maximum speed. It’s about ensuring consistency, scalability, and responsiveness in real-world conditions.

For users, this means assessing performance from various angles:

• Latency—including tail latency (P99/P99.9), which directly affects user experience.
• Throughput—or how much work can be accomplished over a specific period.
• Scalability—the ability to maintain performance levels as demand rises.
• Consistency—ensuring performance doesn’t fluctuate unpredictably under pressure.

Another crucial aspect is time-to-performance: how rapidly infrastructure can be set up, scaled, or restored. The speed at which you can adapt to changes often matters just as much as your system’s runtime.

Azure IaaS integrates these components, aligning compute, storage, and networking to suit specific workloads. This coordination delivers performance as a synergistic system rather than a collection of separate parts.

Boosting AI workloads with system-level performance

AI workloads present some of the toughest performance demands. Training and inference paths need massive parallel compute, high-throughput data access, and quick communication between distributed elements.

In these instances, performance relies on the weakest link. Azure tackles this by optimising the entire data journey.

Enhancing compute efficiency through platform acceleration

Azure Boost enhances VM performance by transferring storage and networking tasks away from the host CPU to dedicated hardware and software, reducing hypervisor overhead. This frees up compute resources for model training and inference, improving both throughput and latency uniformity.

High-throughput storage for uninterrupted data access

AI workloads rely on ongoing access to extensive datasets. Azure’s storage solutions are tailored to provide maintained IO performance, ensuring your compute resources remain active rather than idling for data. Services like Azure Blob Storage and ADLS deliver the high-throughput, low-latency, and highly scalable data foundations that AI workloads require, facilitating rapid ingestion and retrieval of sizeable datasets for training and inference. Their optimised parallel data access and seamless compatibility with AI tools aid in maximising compute usage and eliminating bottlenecks.

Quick, high-bandwidth networking

Distributed training demands speedy communication among nodes. Azure’s networking services, such as Azure ExpressRoute, facilitate quick data movement across clusters, reducing sync delays and boosting overall training efficiency, preventing compute resources from remaining idle.

Together, these features ensure that advancements in compute are not hindered by storage or networking hold-ups. This empowers organisations to process larger data volumes and train models faster without overburdening their infrastructure.

Scaling cloud-native applications without losing performance

Cloud-native applications bring different performance challenges. Instead of fixed workloads, they need to cope with unpredictable demand, scaling up and down dynamically while ensuring responsiveness.

Azure Kubernetes Service (AKS) provides a robust foundation for this flexibility, allowing workloads to scale across nodes. However, just scaling compute isn’t adequate; stateful services also need to scale efficiently.

This is where Azure’s integrated approach truly shines.

Dynamic, high-performance storage for Kubernetes

Azure Container Storage allows AKS workloads to use local NVMe disks through Kubernetes-native provisioning, eliminating manual disk set-up while providing sub-millisecond latency and high IOPS for stateful services.

Ready-to-go data platforms on Kubernetes

With tools like CloudNativePG, companies can run PostgreSQL and other databases directly on AKS, complete with built-in high availability, failover, and backup features, all without compromising performance. Adding flexible data access across file and object storage enhances this foundation, allowing applications to utilise the most suitable storage interface, simplifying data movement and recovery across environments.

Low-latency service communication

Microservices architectures rely on frequent communication between components. Leveraging eBPF host routing within Cilium and Advanced Container Networking Services, we can optimise data flow, enhancing efficiency by reducing latency and boosting throughput, ensuring seamless communication across extensive microservices setups. This keeps inter-service communication quick and consistent, thus averting latency from becoming a bottleneck.

The outcome is a platform that supports both stateless and stateful workloads in dynamic scaling while maintaining peak performance. Furthermore, since resources can be provisioned and scaled as needed, organisations enjoy enhanced cost efficiency, paying only for what they actually use, while keeping applications responsive.

Maintaining performance for critical business systems

For crucial workloads, such as enterprise databases, SAP environments, and transactional systems, it’s not just speed that matters but also predictability and reliability.

These systems require consistent performance under sustained loads, often with strict latency and availability expectations. Any variability, even minor, can have significant repercussions for business.

Azure addresses this through precise regulation and platform-level enhancements.

Reliable compute performance

Azure ensures dependable compute performance through specially designed VM architectures, smart placement, and comprehensive orchestration. Virtual Machine Scale Sets (VMSS) automatically distribute and scale workloads across diverse domains, helping to keep performance predictable amidst varying demand. Azure also boosts consistency further with Azure Boost, which transfers virtualization and I/O processing to dedicated hardware, alleviating contention and enhancing efficiency.

Adjustable storage performance

Azure Ultra Disk and Premium SSD v2 allow clients to separately configure capacity, IOPS, and throughput. This separation helps ensure storage performance aligns precisely with workload demands, steering clear of both underperformance and needless costs.

In addition to customisable block storage like Ultra Disk and Premium SSD v2, Azure offers highly durable object and file storage services—such as Azure Blob Storage and Azure Files—that ensure geo-redundancy and long-term data protection for unstructured data while enhancing performance with enterprise-grade reliability.

Consistent, low-latency networking

Reliable communication between application tiers is vital for transactional systems. Azure’s networking framework guarantees that latency remains low and consistent across settings thanks to features such as Accelerated Networking, speeding up traffic flow by bypassing virtual switch paths, and proximity placement groups that keep latency-sensitive resources physically nearby in the data centre. Coupled with Azure Boost, which allocates networking tasks to dedicated hardware, these capabilities facilitate rapid, consistent data movements, helping to sustain application performance, even at scale.

Rapid recovery

Recovery speed also forms part of performance. Instant Access Snapshots make it possible to restore disks immediately—without the wait for data to hydrate—reducing downtime and expediting recovery from failures.

Azure Backup’s fast restore features further shorten recovery times, while zone-redundant storage (ZRS) maintains data across availability zones, mitigating localized disruptions.

For broader incidents, Azure Site Recovery manages failover across regions, quickly reinstating workloads. This approach is complemented by Azure Disk incremental snapshots that capture only modified data, lowering recovery point objectives (RPO) with minimal overhead, enabling faster, more efficient restoration across various scenarios.

This combination ensures that performance is upheld during regular operations and high-demand periods, as well as recovery situations—where it is most crucial.

Performance as a coordinated system

Between AI, cloud-native, and critical business workloads, a clear pattern develops: performance isn’t just about optimising one single part in isolation.

Performance relies on the effective interaction between compute, storage, and networking tailored to the workload in question.

This synergy helps minimise bottlenecks while ensuring that enhancements in one component are strengthened by capabilities in others. It also simplifies management, letting teams focus on designing workloads and achieving business outcomes rather than fine-tuning the infrastructure.

Practical advice for optimising your workload

While Azure lays a strong foundation, reaching optimal performance still requires aligning your infrastructure with your unique workload needs:

• For AI workloads, aim for balanced throughput across compute, storage, and networking to avoid idle resources and maximise efficiency.
• For cloud-native applications, design for horizontal scaling and consider Kubernetes-native storage to maintain stateful service performance.
• For business-critical systems, prioritise consistency and predictability, using tunable storage and optimised compute to achieve rigorous performance expectations.
• In all cases, examine performance holistically and use platform capabilities to reduce overhead and simplify optimisation.

Establishing a performance-focused foundation

Performance impacts all facets of your application—from user experience to operational efficiency and the potential for innovation.

By merging compute, storage, and networking into a unified platform, Azure empowers companies to deliver excellent performance across their most demanding workloads—without the hassle of managing each layer separately.

Find out how Azure boosts performance across AI, cloud-native, and business critical workloads.

For in-depth resources, visit the Azure IaaS Resource Center for tutorials, best practices, and guidelines on compute, storage, and networking to help you design and manage a resilient infrastructure with confidence.

Have you caught up on the latest posts in the Azure IaaS series?