mpi-stage: High-Performance File Distribution for HPC Clusters
When running containerized workloads on HPC clusters, one of the first problems you hit is getting container images onto the nodes quickly and repeatably. A .sqsh is a Squashfs image (commonly used by container runtimes on HPC). In some environments you can run a Squashfs image directly from shared storage, but at scale that often turns the shared filesystem into a hot spot. Copying the image to local NVMe keeps startup time predictable and avoids hundreds of nodes hammering the same source during job launch. In this post, I'll introduce mpi-stage, a lightweight tool that uses MPI broadcasts to distribute large files across cluster nodes at speeds that can saturate the backend network. The Problem: Staging Files at Scale On an Azure CycleCloud Workspace for Slurm cluster with GB300 GPU nodes, I needed to stage a large Squashfs container image from shared storage onto each node's local NVMe storage before launching training jobs. At small scale you can often get away with ad-hoc copies, but once hundreds of nodes are all trying to read the same source file, the shared source filesystem quickly becomes the bottleneck. I tried several approaches: Attempt 1: Slurm's sbcast Slurm's built-in sbcast seemed like the natural choice. In my quick testing it was slower than I wanted, and the overwrite/skip-existing behavior didn't match the "fast no-op if already present" workflow I was after. I didn't spend much time exploring all the configuration options before moving on. Attempt 2: Shell Script Fan-Out I wrote a shell script using a tree-based fan-out approach: copy to N nodes, then each of those copies to N more, and so on. This worked and scaled reasonably, but had some drawbacks: Multiple stages: The script required orchestrating multiple rounds of copy commands, adding complexity Source filesystem stress: Even with fan-out, the initial copies still hit the source filesystem simultaneously — a fan-out of 4 meant 4 nodes competing for source bandwidth Frontend network: Copies went over the Ethernet network by default — I could have configured IPoIB, but that added more setup The Solution: MPI Broadcasts The key insight was that MPI's broadcast primitive (MPI_Bcast) is specifically optimized for one-to-many data distribution. Modern MPI implementations like HPC-X use tree-based algorithms that efficiently utilize the high-bandwidth, low-latency InfiniBand network. With mpi-stage: Single source read: Only one node reads from the source filesystem Backend network utilization: Data flows over InfiniBand using optimized MPI collectives Intelligent skipping: Nodes that already have the file (verified by size or checksum) skip the copy entirely Combined, this keeps the shared source (NFS, Lustre, blobfuse, etc.) from being hammered by many concurrent readers while still taking full advantage of the backend fabric. How It Works mpi-stage is designed around a simple workflow: The source node reads the file in chunks and streams each chunk via MPI_Bcast. Destination nodes write each chunk to local storage immediately upon receipt. This streaming approach means the entire file never needs to fit in memory — only a small buffer is required. Key Features Pre-copy Validation Before any data is transferred, each node checks if the destination file already exists and matches the source. You can choose between: Size check (default): Fast comparison of file sizes—sufficient for most use cases Checksum: Stronger validation, but requires reading the full file and is therefore slower If all nodes already have the correct file, mpi-stage completes in milliseconds with no data transfer. Double-Buffered Transfers The implementation uses double-buffered, chunked transfers to overlap network communication with disk I/O. While one buffer is being broadcast, the next chunk is being read from the source. Post-copy Validation Optionally verify that all nodes received the file correctly after the copy completes. Single-Writer Per Node The tool enforces one MPI rank per node to prevent filesystem contention and ensure predictable performance. Real-World Performance In one run using 156 GPU nodes, distributing a container image achieved approximately 3 GB/s effective distribution rate (file_size/time), completing in just over 5 seconds: [0] Copy required: yes [0] Starting copy phase (source writes: yes) [0] Copy complete, Bandwidth: 3007.14 MB/s [0] Post-validation complete [0] Timings (s): Topology check: 5.22463 Source metadata: 0.00803746 Pre-validation: 0.0046786 Copy phase: 5.21189 Post-validation: 2.2944e-05 Total time: 5.2563 Because every node writes the file to its own local NVMe, the cumulative write rate across the cluster is roughly this number times the node count: ~3 GB/s × 156 ≈ ~468 GB/s of total local writes. Workflow: Container Image Distribution The primary use case is distributing Squashfs images to local NVMe before launching containerized workloads. Run mpi-stage as a job step before your main application: #!/bin/bash #SBATCH --job-name=my-training-job #SBATCH --ntasks-per-node=1 #SBATCH --exclusive # Stage the container image srun --mpi=pmix ./mpi_stage \ --source /shared/images/pytorch.sqsh \ --dest /nvme/images/pytorch.sqsh \ --pre-validate size \ --verbose # Run the actual job (from local NVMe - much faster!) srun --container-image=/nvme/images/pytorch.sqsh ... mpi-stage will create the destination directory if it doesn't exist. If your container runtime supports running the image directly from shared storage, you may not strictly need this step—but staging to local NVMe tends to be faster and more predictable at large scale. Because of the pre-validation, you can include this step in every job script without penalty—if the image is already present, it completes in milliseconds. Getting Started git clone https://github.com/edwardsp/mpi-stage.git cd mpi-stage make For detailed usage and options, see the README. Summary mpi-stage started as a solution to a very specific problem—staging large container images efficiently across a large GPU cluster—but the same pattern may be useful in other scenarios where many nodes need the same large file. By using MPI broadcasts, only a single node reads from the source filesystem, while data is distributed over the backend network using optimized collectives. In practice, this can significantly reduce load on shared filesystems and cloud-backed mounts, such as Azure Blob Storage accessed via blobfuse2, where hundreds of concurrent readers can otherwise become a bottleneck. While container images were the initial focus, this approach could also be applied to staging training datasets, distributing model checkpoints or pretrained weights, or copying large binaries to local NVMe before a job starts. Anywhere that a “many nodes, same file” pattern exists is a potential fit. If you're running large-scale containerized workloads on Azure HPC infrastructure, give it a try. If you use mpi-stage in other workflows, I'd love to hear what worked (and what didn't). Feedback and contributions are welcome. Have questions or feedback? Leave a comment below or open an issue on GitHub.Monitoring HPC & AI Workloads on Azure H/N VMs Using Telegraf and Azure Monitor (GPU & InfiniBand)
As HPC & AI workloads continue to scale in complexity and performance demands, ensuring visibility into the underlying infrastructure becomes critical. This guide presents an essential monitoring solution for AI infrastructure deployed on Azure RDMA-enabled virtual machines (VMs), focusing on NVIDIA GPUs and Mellanox InfiniBand devices. By leveraging the Telegraf agent and Azure Monitor, this setup enables real-time collection and visualization of key hardware metrics, including GPU utilization, GPU memory usage, InfiniBand port errors, and link flaps. It provides operational insights vital for debugging, performance tuning, and capacity planning in high-performance AI environments. In this blog, we'll walk through the process of configuring Telegraf to collect and send GPU and InfiniBand monitoring metrics to Azure Monitor. This end-to-end guide covers all the essential steps to enable robust monitoring for NVIDIA GPUs and Mellanox InfiniBand devices, empowering you to track, analyze, and optimize performance across your HPC & AI infrastructure on Azure. DISCLAIMER: This is an unofficial configuration guide and is not supported by Microsoft. Please use it at your own discretion. The setup is provided "as-is" without any warranties, guarantees, or official support. While Azure Monitor offers robust monitoring capabilities for CPU, memory, storage, and networking, it does not natively support GPU or InfiniBand metrics for Azure H- or N-series VMs. To monitor GPU and InfiniBand performance, additional configuration using third-party tools—such as Telegraf—is required. As of the time of writing, Azure Monitor does not include built-in support for these metrics without external integrations. 🔔 Update: Supported Monitoring Option Now Available Update (December 2025): At the time this guide was written, monitoring InfiniBand (IB) and GPU metrics on Azure H-series and N-series VMs required a largely unofficial approach using Telegraf and Azure Monitor. Microsoft has since introduced a supported solution: Azure Managed Prometheus on VM / VM Scale Sets (VMSS), currently available in private preview. This new capability provides a native, managed Prometheus experience for collecting infrastructure and accelerator metrics directly from VMs and VMSS. It significantly simplifies deployment, lifecycle management, and long-term support compared to custom Telegraf-based setups. For new deployments, customers are encouraged to evaluate Azure Managed Prometheus on VM / VMSS as the preferred and supported approach for HPC and AI workload monitoring. Official announcement: Private Preview: Azure Managed Prometheus on VM / VMSS Step 1: Making changes in Azure for sending GPU and IB metrics from Telegraf agents to Azure monitor from VM or VMSS. Register the microsoft.insights resource provider in your Azure subscription. Refer: Resource providers and resource types - Azure Resource Manager | Microsoft Learn Step 2: Enable Managed Service Identities to authenticate an Azure VM or Azure VMSS. In the example we are using Managed Identity for authentication. You can also use User Managed Identities or Service Principle to authenticate the VM. Refer: telegraf/plugins/outputs/azure_monitor at release-1.15 · influxdata/telegraf (github.com) Step 3: Set Up the Telegraf Agent Inside the VM or VMSS to Send Data to Azure Monitor In this example, I'll use an Azure Standard_ND96asr_v4 VM with the Ubuntu-HPC 2204 image to configure the environment for VMSS. The Ubuntu-HPC 2204 image comes with pre-installed NVIDIA GPU drivers, CUDA, and InfiniBand drivers. If you opt for a different image, ensure that you manually install the necessary GPU drivers, CUDA toolkit, and InfiniBand driver. Next, download and run the gpu-ib-mon_setup.sh script to install the Telegraf agent on Ubuntu 22.04. This script will also configure the NVIDIA SMI input plugin and InfiniBand Input Plugin, along with setting up the Telegraf configuration to send data to Azure Monitor. Note: The gpu-ib-mon_setup.sh script is currently supported and tested only on Ubuntu 22.04. Please read the InfiniBand counter collected by Telegraf - https://enterprise-support.nvidia.com/s/article/understanding-mlx5-linux-counters-and-status-parameters Run the following commands: wget https://raw.githubusercontent.com/vinil-v/gpu-ib-monitoring/refs/heads/main/scripts/gpu-ib-mon_setup.sh -O gpu-ib-mon_setup.sh chmod +x gpu-ib-mon_setup.sh ./gpu-ib-mon_setup.sh Test the Telegraf configuration by executing the following command: sudo telegraf --config /etc/telegraf/telegraf.conf --test Step 4: Creating Dashboards in Azure Monitor to Check NVIDIA GPU and InfiniBand Usage Telegraf includes an output plugin specifically designed for Azure Monitor, allowing custom metrics to be sent directly to the platform. Since Azure Monitor supports a metric resolution of one minute, the Telegraf output plugin aggregates metrics into one-minute intervals and sends them to Azure Monitor at each flush cycle. Metrics from each Telegraf input plugin are stored in a separate Azure Monitor namespace, typically prefixed with Telegraf/ for easy identification. To visualize NVIDIA GPU usage, go to the Metrics section in the Azure portal: Set the scope to your VM. Choose the Metric Namespace as Telegraf/nvidia-smi. From there, you can select and display various GPU metrics such as utilization, memory usage, temperature, and more. In example we are using GPU memory_used metrics. Use filters and splits to analyze data across multiple GPUs or over time. To monitor InfiniBand performance, repeat the same process: In the Metrics section, set the scope to your VM. Select the Metric Namespace as Telegraf/infiniband. You can visualize metrics such as port status, data transmitted/received, and error counters. In this example, we are using a Link Flap Metrics to check the InfiniBand link flaps. Use filters to break down the data by port or metric type for deeper insights. Link_downed Metric Note: The link_downed metric with Aggregation: Count is returning incorrect values. We can use Max, Min values. Port_rcv_data metrics Creating custom dashboards in Azure Monitor with both Telegraf/nvidia-smi and Telegraf/infiniband namespaces allows for unified visibility into GPU and InfiniBand. Testing InfiniBand and GPU Usage If you're testing GPU metrics and need a reliable way to simulate multi-GPU workloads—especially over InfiniBand—here’s a straightforward solution using the NCCL benchmark suite. This method is ideal for verifying GPU and network monitoring setups. NCCL Benchmark and OpenMPI is part of the Ubuntu HPC 22.04 image. Update the variable according to your environment. Update the hostfile with the hostname. module load mpi/hpcx-v2.13.1 export CUDA_VISIBLE_DEVICES=2,3,0,1,6,7,4,5 mpirun -np 16 --map-by ppr:8:node -hostfile hostfile \ -mca coll_hcoll_enable 0 --bind-to numa \ -x NCCL_IB_PCI_RELAXED_ORDERING=1 \ -x LD_LIBRARY_PATH=/usr/local/nccl-rdma-sharp-plugins/lib:$LD_LIBRARY_PATH \ -x CUDA_DEVICE_ORDER=PCI_BUS_ID \ -x NCCL_SOCKET_IFNAME=eth0 \ -x NCCL_TOPO_FILE=/opt/microsoft/ndv4-topo.xml \ -x NCCL_DEBUG=WARN \ /opt/nccl-tests/build/all_reduce_perf -b 8 -e 8G -f 2 -g 1 -c 1 Alternate: GPU Load Simulation Using TensorFlow If you're looking for a more application-like load (e.g., distributed training), I’ve prepared a script that sets up a multi-GPU TensorFlow training environment using Anaconda. This is a great way to simulate real-world GPU workloads and validate your monitoring pipelines. To get started, run the following: wget -q https://raw.githubusercontent.com/vinil-v/gpu-monitoring/refs/heads/main/scripts/gpu_test_program.sh -O gpu_test_program.sh chmod +x gpu_test_program.sh ./gpu_test_program.sh With either method NCCL benchmarks or TensorFlow training you’ll be able to simulate realistic GPU usage and validate your GPU and InfiniBand monitoring setup with confidence. Happy testing! References: Ubuntu HPC on Azure ND A100 v4-series GPU VM Sizes Telegraf Azure Monitor Output Plugin (v1.15) Telegraf NVIDIA SMI Input Plugin (v1.15) Telegraf InfiniBand Input Plugin Documentation
Performance and Scalability of Azure HBv5-series Virtual Machines
Azure HBv5-series virtual machines (VMs) for CPU-based high performance computing (HPC) are now Generally Available. This blog provides in-depth information about the technical underpinnings, performance, cost, and management implications of these HPC-optimized VMs. Azure HBv5 VM bring leadership levels of performance, cost optimization, and server (VM) consolidation for a variety of workloads driven by memory performance, such as computational fluid dynamics, weather simulation, geoscience simulations, and finite element analysis. For these applications and compared to HBv4 VMs, previously the highest performance offering for these workloads, HBv5 provides up to : 5x higher performance for CFD workloads with 43% lower costs 3.2x higher performance for weather simulation with 16% lower costs 2.8x higher performance for geoscience workloads at the same costs HBv5-series Technical Overview & VM Sizes Each HBv5 VMs features several new technologies for HPC customers, including: Up to 6.6 TB/s of memory bandwidth (STREAM TRIAD) and 432 GB memory capacity Up to 368 physical cores per VM (user configurable) with custom AMD EPYC CPUs, Zen4 microarchitecture (SMT disabled) Base clock of 3.5 GHz (~1 GHz higher than other 96-core EPYC CPUs), and Boost clock of 4 GHz across all cores 800 Gb/s NVIDIA Quantum-2 InfiniBand (4 x 200 Gb/s CX-7) (~2x higher HBv4 VMs) 180 Gb/s Azure Accelerated Networking (~2.2 higher than HBv4 VMs) 15 TB local NVMe SSD with up to 50 GB/s (read) and 30 GB/s (write) of bandwidth (~4x higher than HBv4 VMs) The highlight feature of HBv5 VMs is their use of high-bandwidth memory (HBM). HBv5 VMs utilize a custom AMD CPU that increases memory bandwidth by ~9x v. dual-socket 4 th Gen EPYC (Zen4, “Genoa”) server platforms, and ~7x v. dual-socket EPYC (Zen5, “Turin”) server platforms, respectively. HBv5 delivers similar levels of memory bandwidth improvement compared to the highest end alternatives from the Intel Xeon and ARM CPU ecosystems. HBv5-series VMs are available in the following sizes with specifications as shown below. Just like existing H-series VMs, HBv5-series includes constrained cores VM sizes, enabling customers to optimize their VM dimensions for a variety of scenarios: ISV licensing constraining a job to a targeted number of cores Maximum-performance-per-VM or maximum performance per core Minimum RAM/core (1.2 GB, suitable for strong scaling workloads) to maximum memory per core (9 GB, suitable for large datasets and weak scaling workloads Table 1: Technical specifications of HBv5-series VMs Note: Maximum clock frequencies (FMAX) are based product specifications of the AMD EPYC 9V64H processor. Experienced clock frequencies by a customer are a function of a variety of factors, including but not limited to the arithmetic intensity (SIMD) and parallelism of an application. For more information see official documentation for HBv5-series VMs Microbenchmark Performance This section focuses on microbenchmarks that characterize performance of the memory subsystem, compute capabilities, and InfiniBand network of HBv5 VMs. Memory & Compute Performance To capture synthetic performance, we ran the following industry standard benchmarks: STREAM – memory bandwidth High Performance Conjugate Gradient (HPCG) – sparse linear algebra High Performance Linpack (HPL)– dense linear algebra Absolute results and comparisons to HBv4 VMs are shown in Table 2, below: Table 2: Results of HBv5 running the STREAM, HPCG, and HPL benchmarks. Note: STREAM was run with the following CLI parameters: OMP_NUM_THREADS=368 OMP_PROC_BIND=true OMP_PLACES=cores ./amd_zen_stream STREAM data size: 2621440000 bytes InfiniBand Networking Performance Each HBv5-series VM is equipped with four NVIDIA Quantum-2 network interface cards (NICs), each operating at 200 Gb/s for an aggregate bandwidth of 800 Gb/s per VM (node). We ran the industry standard IB perftests based on OSU benchmarks test across two (2) HBv5-series VMs, as depicted in the results shown in Figures 3-5, below: Note: all results below are for a single 200 Gb/s (uni-directional) link only. At a VM level, all bandwidth results below are 4x higher as there are four (4) InfiniBand links per HBv5 server. Unidirectional bandwidth: numactl -c 0 ib_send_bw -aF -q 2 Figure 1: results showing 99% achieved uni-directional bandwidth v. theoretical peak. Bi-directional bandwidth: numactl -c 0 ib_send_bw -aF -q 2 -b Figure 2: results showing 99% achieved bi-directional bandwidth v. theoretical peak. Latency: Figure 3: results measuring as low as 1.25 microsecond latencies among HBv5 VMs. Latencies experienced by users will depend on message sizes employed by applications. Application Performance, Cost/Performance, and Server (VM) Consolidation This section focuses on characterizing HBv5-series VMs when running common, real-world HPC applications with an emphasis on those known to be meaningfully bound by memory performance as that is the focus of the HB-series family. We characterize HBv5 below in three (3) ways of high relevance to customer interests: Performance (“how much faster can it do the work”) Cost/Performance (“how much can it reduce the costs to complete the work”) Fleet consolidation (“how much can a customer simplify the size and scale of compute fleet management while still being able to the work”) Where possible, we have included comparisons to other Azure HPC VMs, including: Azure HBv4/HX series with 176 physical cores of 4 th Gen AMD EPYC CPUs with 3D V-Cache (“Genoa-X”) (HBv4 specifications, HX specifications) Azure HBv3 with 120 physical cores of 3 rd Gen AMD EPYC CPUs with 3D V-Cache (“Milan-X”) (HBv3 specifications) Azure HBv2 with 120 physical cores of 2 nd Gen AMD EPYC CPUs (“Rome”) processors (full specifications) Unless otherwise noted, all tests shown below were performed with: Alma Linux 8.10 (image URN : almalinux:almalinux-hpc:8_10-hpc-gen2:latest) for scaling ( image URN: almalinux:almalinux-hpc:8_6-hpc-gen2:latest) NVIDIA HPC-X MPI Further, all Cost/Performance comparisons leverage pricing rate info from list price, Pay-As-You-Go (PAYG) information found on Azure Linux Virtual Machines Pricing. Absolute costs will be a function of a customer’s workload, model, and consumption (PAYG v. Reserved Instance, etc.) approach. That said, the relative cost/performance comparisons illustrated below should hold for the workload and model combinations shown below, regardless of the consumption approach. Computational Fluid Dynamics (CFD) OpenFOAM – version 2306 with 100M Cell Motorbike case Figure 4: HBv5 v. HBv4 on on OpenFOAM with the Motorbike 100M cell case HBv5 VMs provide a 4.8x performance increase over HBv4 VMs. Figure 5: The cost to complete the OpenFOAM Motorbike 100M case is just 57% of what it costs to complete the same case on HBv4. Above, we can see that for customers running OpenFOAM cases similar to the size and complexity of the 100M cell Motorbike problem, organizations can consolidate their server (VM) deployments by approximately a factor of five (5). Palabos – version 1.01 with 3D Cavity, 1001 x 1001 x 1001 cells case Figure 6: On Palabos, a Lattice Boltzmann solver using a streaming memory access pattern, HBv5 VMs provide a 4.4x performance increase over HBv4 VMs. Figure 7: The cost to complete the Palabos 3D Cavity case is just 62% of what it costs to complete the same case on HBv4. Above, we can see that for customers running Palabos with cases similar to the size and complexity of the 100M cell Motorbike problem, organizations can consolidate their server (VM) deployments by approximately a factor of ~4.5. Ansys Fluent – version 2025 R2 with F1 Racecar 140M case Figure 8: On ANSYS Fluent HBv5 VMs provide a 3.4x performance increase over HBv4 VMs. Figure 9: The cost to complete the ANSYS Fluent F1 racecar 140M case is just 81% of what it costs to complete the same case on HBv4. Above, we can see that for customers running ANSYS Fluent with cases similar to the size and complexity of the 140M cell F1 Racecar problem, organizations can consolidate their server (VM) deployments by approximately a factor of ~3.5. Siemens Star-CCM+ - version 17.04.005 with AeroSUV Steady Coupled 106M case Figure 10: On Star-CCM+, HBv5 VMs provide a 3.4x performance increase over HBv4 VMs. Figure 11: The cost to complete the Siemens Star-CCM+ANSYS Fluent F1 racecar 140M case is just 81% of what it costs to complete the same case on HBv4. Above, we can see that for customers running Star-CCM+ with cases similar to the size and complexity of the 106M cell AeroSUV Steady Coupled, organizations can consolidate their server (VM) deployments by approximately a factor of ~3.5. Weather Modeling WRF – version 4.2.2 with CONUS 2.5KM case Figure 12: On WRF, HBv5 VMs provide a 3.27x performance increase over HBv4 VMs. Figure 13: The cost to complete the WRF Conus 2.5KM case is just 84% of what it costs to complete the same case on HBv4. Above, we can see that for customers running WRF with cases similar to the size and complexity of the 2.5km CONUS, organizations can consolidate their server (VM) deployments by approximately a factor of ~3. Energy Research Devito – version 4.8.7 with Acoustic Forward case Figure 14: On Devito, HBv5 VMs provide a 3.27x performance increase over HBv4 VMs. Figure 15: The cost to complete the Devito Acoustic Forward OP case is equivalent to what it costs to complete the same case on HBv4. Above, we can see that for customers running Devito with cases similar to the size and complexity of the Acoustic Forward OP, organizations can consolidate their server (VM) deployments by approximately a factor of ~3. Molecular Dynamics NAMD - version 2.15a2 with STMV 20M case Figure 16: On NAMD, HBv5 VMs provide a 2.18x performance increase over HBv4 VMs. Figure 17: The cost to complete the NAMD STMV 20M case is 26% higher on HBv5 than what it costs to complete the same case on HBv4 Above, we can see that for customers running NAMD with cases similar to the size and complexity of the STMV 20M case, organizations can consolidate their server (VM) deployments by approximately a factor of ~2. Notably, NAMD is a compute bound case, rather than memory performance bound. We include it here to illustrate that not all workloads are fit for purpose with HBv5. This latest Azure HPC VM is the fastest at this workload on the Microsoft Cloud, but does not benefit substantially from HBv5’s premium levels of memory bandwidth. NAMD would instead perform more cost efficiently with a CPU that supports AVX512 instructions natively or, much better still, a modern GPU. Scalability of HBv5-series VMs Weak Scaling Weak scaling measures how well a parallel application or system performs when both the number of processing elements and the problem size increase proportionally, so that the workload per processor remains constant. Weak scaling cases are often employed when time-to-solution is fixed (e.g. it is acceptable to solve a problem within a specified period) but a user desires a simulation to be of a higher fidelity or resolution. A common example is operational weather forecasting. To illustrate weak scaling on HBv5 VMs, we ran Palabos with the same 3D cavity problem as shown earlier: Figure 18: On Palabos with the 3D Cavity model, HBv5 scales linearly as the 3D cavity size is proportionately increased. Strong Scaling Strong scaling is characterized by the efficiency with which execution time is reduced as the number of processor elements (CPUs, GPUs, etc.) is increased, while the problem size remains kept constant. Strong scaling cases are often employed when the fidelity or resolution of the simulation is acceptable, but a user requires faster time to completion. A common example is product engineering validation when an organization wants to bring a product to market faster but must complete a broad range of validation and verification scenarios before doing so. To illustrate Strong scaling on HBv5 VMs, we ran NAMD with two different problems, each intended to illustrate the how expectations for strong scaling efficiency change depending on problem size and the ordering of computation v. communication in distributed memory workloads. First, let us examine NAMD with the 20M STMV benchmark Figure 19: Strong scaling on HBv5 with NAMD STMV 20M cell case As illustrated above, for strong scaling cases for which the compute time is continuously reduced (by leveraging more and more processor elements) but communication time remains constant, scaling efficiency will only stay high for so long. That principle is well-represented by the STMV 20m case, for which parallel efficiency remains linear (i.e. cost/job remains flat) at two (2) nodes but degrades after that. This is because while compute is being sped up, the MPI time remains relatively flat. As such, the relatively static MPI time comes to dominate end-to-end wall clock time as VM scaling increases. Said another way, HBv5 features so much compute performance that even for a moderate-sized problem like STMV 20M scaling the infrastructure can only take performance so far and cost/job will begin to increase. If we examine HBv5 against the 210M cell case, however, with 10.5x as many elements to compute as its 20M case sibling, the scaling efficiency story changes significantly. Figure 19: On NAMD with the STMV 210M cell case, HBv5 scales linearly out to 32 VMs (or more than 11,000 CPU cores). As illustrated above, larger cases with significant compute requirements will continue to scale efficiently with larger amounts of HBv5 infrastructure. While MPI time remains relatively flat for this case (as is the case with the smaller STMV 20M case), the compute demands remain the dominant fraction of end-to-end wall clock time. As such, HBv5 scales these problems with very high levels of efficiency and in doing so job costs to the user remain flat despite up to 8x as many VMs being leveraged compared to the four (4) VM baseline. The key takeaways for strong scaling scenarios are two-fold. First, users should run scaling tests with their applications and models to find a sweet spot of faster performance with constant job costs. This will depend heavily on model size. Second, as new and very high end compute platforms like HBv5 emerge that accelerate compute time, application developers will need to find ways reduce wall clock times bottlenecking on communication (MPI) time. Recommended approaches include using fewer MPI processes and, ideally, restructuring applications to overlap communication with compute phases.
Breaking the Million-Token Barrier: The Technical Achievement of Azure ND GB300 v6
Azure ND GB300 v6 Virtual Machines with NVIDIA GB300 NVL72 rack-scale systems achieve unprecedented performance of 1,100,000 tokens/s on Llama2 70B Inference, beating the previous Azure ND GB200 v6 record of 865,000 tokens/s by 27%.
A Quick Guide to Benchmarking AI models on Azure: Llama 405B and 70B with MLPerf Inference v5.1
by Mark Gitau (Software Engineer) Introduction For the MLPerf Inference v5.1 submission, Azure shared performance results on the new ND GB200 v6 virtual machines. A single ND GB200 v6 VM on Azure is powered by two NVIDIA Grace CPUs and four NVIDIA Blackwell B200 GPUs. This document highlights Azure’s MLPerf Inference v5.1 results and outlines the steps to run these benchmarks on Azure. These MLPerf™ benchmark results demonstrate Azure’s commitment to providing our customers with the latest GPU offerings of the highest quality. Highlights from MLPerf Inference v5.1 benchmark results include: Azure had the highest Llama 2 70B Offline submission with 52,000 tokens/s on a single ND GB200 v6 virtual machine. This corresponds to an 8% increase on single node performance since our record which would correspond to 937,098 tokens/s on a full NVL72 rack. Azure results for Llama 3.1 405B are at par with the best submitters (1% difference), cloud and on-premises, with 847 tokens/s. How to replicate the results in Azure Pre-requisites: ND GB200 v6-series (single node): Deploy and set up a virtual machine on Azure Set up the environment First, we need to export the path to the directory where we will perform the benchmarks. For ND GB200 v6-series (single node), create a directory called mlperf in /mnt/nvme Set mlperf scratch space: export MLPERF_SCRATCH_PATH=/mnt/nvme/mlperf Clone the MLPerf repository inside the scratch path: git clone https://github.com/mlcommons/inference_results_v5.1.git Then create empty directories in your scratch space to house the data: mkdir $MLPERF_SCRATCH_PATH/data $MLPERF_SCRATCH_PATH/models $MLPERF_SCRATCH_PATH/preprocessed_data Download the models & datasets Download the models inside the models directory you created in the previous step. This will take a while because the weights are large. Llama 2 70B model Llama 3.1 405B model Download the preprocessed datasets for both models Llama 2 70B datasets Llama 3.1 405B datasets Prepare the datasets for Llama 2 70B: inference_results_v5.1/closed/Azure/code/llama2-70b/tensorrt at main · mlcommons/inference_results_v5.1 prepare the datasets for Llama 3.1 405B: inference_results_v5.1/closed/Azure/code/llama3.1-405b/tensorrt at main · mlcommons/inference_results_v5.1 Build & launch MLPerf container Export the Submitter and System name: export SUBMITTER=Azure SYSTEM_NAME=ND_GB200_v6 Enter the container by entering the closed/Azure directory and running: make prebuild Inside the container, run make build Build engines & run benchmarks Make sure you are still in the closed/Azure directory of the MLPerf repository To build the engines for both Llama 3.1 405B and Llama 2 70B: make generate_engines RUN_ARGS="--benchmarks=llama2-70b,llama3.1-405b --scenarios=offline,server" To run the benchmarks for both Llama 3.1 405B and Llama 2 70B: make run_harness RUN_ARGS=="--benchmarks=llama2-70b,llama3.1-405b --scenarios=offline,server" MLPerf from MLCommons® MLCommons® is an open engineering consortium of AI leaders from academia, research, and industry where the mission is to “build fair and useful benchmarks” that provide unbiased evaluations of training and inference performance for hardware, software, and services—all conducted under predetermined conditions. MLPerf™ Inference benchmarks consist of compute-intensive AI workloads that simulate realistic usage of the systems, making the results very influential in technology management’s buying decisions.Azure’s ND GB200 v6 Delivers Record Performance for Inference Workloads
Achieving peak AI performance requires both cutting-edge hardware and a finely optimized infrastructure. Azure’s ND GB200 v6 Virtual Machines, accelerated by the NVIDIA GB200 Blackwell GPUs, have already demonstrated world record performance of 865,000 tokens/s for inferencing on the industry standard LLAMA2 70B
Performance analysis of DeepSeek R1 AI Inference using vLLM on ND-H100-v5
Introduction The DeepSeek R1 model represents a new frontier in large-scale reasoning for AI applications. Designed to tackle complex inference tasks, R1 pushes the boundaries of what’s possible—but not without significant infrastructure demands. To deploy DeepSeek R1 effectively in an inference service like vLLM, high-performance hardware is essential. Specifically, the model requires two Azure ND_H100_v5 nodes, each equipped with 8 NVIDIA H100 GPUs, totaling 16 H100s. These nodes are interconnected via InfiniBand and NVLink, ensuring the bandwidth and latency characteristics necessary to support the model’s massive memory footprint and parallel processing needs. In this post, we’ll present inference benchmark results for DeepSeek R1, measuring performance across GPU utilization, memory throughput, and interconnect efficiency. While R1 excels in reasoning tasks, it’s important to recognize that such models are not universally optimal. For many general-purpose AI applications, smaller models like Llama 3.1 8B offer a compelling alternative, delivering sufficient accuracy and performance at a fraction of the cost. We explore the performance characteristics of DeepSeek R1 and help you decide when a large reasoning model is worth the investment—and when a leaner solution might be the better choice. Benchmark environment 2 ND_H100_v5 nodes (16 H100) were required to load the DeepSeek R1 model into GPU memory. vLLM was used to set-up the online inference server using a single API server endpoint. vLLM bench (specifically hugging face AI-MO/aimo-validation-aime) was used to run the inference benchmark. Detailed specifications are available in the tables below. Results Reasoning model’s additional chain of thought tokens We show that DeepSeek R1 generates many additional tokens (for chain of thought), which helps reasoning models answer complex logical questions. In this example the DeepSeek R1 generates 1162 completion tokens, but Llama 3.1 8B only generates 37. This example also demonstrates that while non-reasoning models may struggle with this seemingly simple logical question, reasoning models like DeepSeek R1 can solve it—though they require significantly more intermediate tokens to do so. #!/bin/bash curl http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "deepseek-ai/DeepSeek-R1", "messages": [ {"role": "user", "content": "9.11 and 9.8, which is greater? Explain your reasoning"} ] }' Note: The REST API json output is formatted to make the output more readable. (./run_restapi_query.sh 2> /dev/null | python3 ./convert_json_2_txt.py >& run_query.out_$$) Reasoning: Okay, so I need to figure out whether 9.11 is greater than 9.8 or if 9.8 is greater. Hmm, let me start by looking at both numbers. They both have a 9 in the units place, so that part is equal. Now, I need to compare the decimal parts: .11 and .8. Wait, but decimals can be tricky because the places matter. The first number is 9.11, which has a 1 in the tenths place and a 1 in the hundredths place. The second number is 9.8, which is the same as 9.80 when you add a zero to the hundredths place for comparison. So, 9.80 versus 9.11. Now, looking at the tenths place: 8 is greater than 1. Therefore, 9.80 should be greater than 9.11. But let me double-check because sometimes when numbers have different decimal lengths, it's easy to make a mistake. If I convert both to fractions, maybe that will help. 9.11 is the same as 9 + 0.11, and 0.11 is 11/100. On the other hand, 9.8 is 9 + 0.8, which is 9 + 8/10. To compare 8/10 and 11/100, I can convert them to have the same denominator. 8/10 is equal to 80/100. Now, 80/100 is much larger than 11/100. So yes, 9.8 is indeed greater than 9.11. Another way to see it is by subtracting the two numbers. If I subtract 9.11 from 9.8, what do I get? Let's see: 9.8 - 9.11. Since 9.8 is the same as 9.80, subtracting 9.11 gives 0.69. Since the result is positive, that means 9.8 is larger. Wait, maybe I can also think about money to make it clearer. If I have $9.11 and you have $9.80, which is more? $9.80 is $9 and 80 cents, whereas $9.11 is $9 and 11 cents. Definitely, 80 cents is more than 11 cents, so $9.80 is more. That helps visualize it. Alternatively, using a number line. If I imagine a number line with 9.0, 9.1, 9.2,... up to 10.0. 9.11 is just a little past 9.1, while 9.8 is almost at 10.0. So clearly, 9.8 is much further to the right, meaning it's a larger number. I think that covers different methods: decimal place comparison, converting to fractions, subtraction, real-life money examples, and number line visualization. All of these methods point to 9.8 being greater than 9.11. I don't see any contradictions here. Maybe I can also consider writing both numbers with the same number of decimal places to make it even clearer. 9.11 is already two decimal places, and 9.8 can be written as 9.80. Now, comparing 9.80 and 9.11 digit by digit: - The units place: 9 vs. 9 (equal) - Tenths place: 8 vs. 1 (8 is greater) - Since the tenths place of 9.80 is greater, the entire number is larger, regardless of the hundredths place. So yes, that confirms it again. The key idea here is that when comparing decimals, you start from the leftmost digit after the decimal point and move right until you find a difference. In this case, the tenths place already shows that 9.8 is larger. I guess another common mistake could be thinking that because 11 is more than 8, 0.11 is more than 0.8, but that's not correct because the places are different. Tenths versus hundredths. 0.8 is equivalent to 0.80, which is way bigger than 0.11. To prevent that mistake, always align the decimal points and add trailing zeros if necessary. Then compare each digit one by one from left to right. That method avoids confusion. So, all in all, through multiple approaches, 9.8 is definitely greater than 9.11. Answer: To determine which number is greater between **9.11** and **9.8**, follow these steps: **Align the decimal places** by rewriting 9.8 as **9.80**. **Compare digit by digit** from left to right: - **Units place**: Both have **9** (equal). - **Tenths place**: **8** (in 9.80) vs. **1** (in 9.11). Since \(8 > 1\), **9.80 is greater** at this point. - No need to compare hundredths place once a difference is found. **Conclusion**: \(9.8\) (or \(9.80\)) is greater than \(9.11\) because **8 tenths** (\(0.8\)) is larger than **1 tenth + 1 hundredth** (\(0.11\)). \(\boxed{9.8 \text{ is greater}}\) Completion Info: Finish reason: stop Prompt tokens: 21 Completion tokens: 1162 Total tokens: 1183 When running using Llama 3.1 8B (Not reasoning model). Answer: To determine which is greater, we need to compare the two numbers. 9.11 is greater than 9.8 because 9.11 is a larger decimal value. Completion Info: Finish reason: stop Prompt tokens: 51 Completion tokens: 37 Total tokens: 88 Throughput and latency results Cost comparison In this cost analysis we use the ND-H100-v5 and ND-H200-v4 pay as you go pricing in south central US region and the measured total throughput tokens/sec to compute the $/(1K tokens). Note: ND-H200-v5 pricing was estimated at 20% more than ND-H100-v5 pricing. Analysis The DeepSeek R1 is a large, complex reasoning model that is costlier and slower than smaller models. It needs 16 H100 GPUs for FP8 precision and generates many more intermediate tokens in its chain of thought process—about 31 times more than Llama 3.1 8B—but at a much slower rate (~54 times slower). Its latency is also higher, with TTFT and ITL being roughly 6 and 3 times slower, respectively. The DeepSeek R1 model has small intranode and internode network requirements (~14% of available InfiniBand network bandwidth was used, and < 1% of available NVLink bandwidth is used. GPUs with higher memory bandwidth and higher FLOPS would help improve its performance. The cost analysis shows that the cost to generate DeepSeek R1 tokens is ~54 times more expensive than Llama 3.1 8B on the same 16 H100 GPU’s and ~34 times more expensive on 8 H200 GPU’s. DeepSeek R1 model is very capability, but due to its higher TCO it should be only used in specific AI applications that require its strong reasoning abilities. Conclusion The DeepSeek R1 model demonstrates exceptional reasoning capabilities, but its deployment demands substantial infrastructure and incurs high latency and cost. While it excels in generating detailed chains of thought, its throughput and efficiency lag significantly behind smaller models like Llama 3.1 8B. For applications requiring deep logical analysis, DeepSeek R1 is a powerful tool. However, for general-purpose inference tasks, more lightweight models offer better performance and cost-effectiveness. Strategic use of DeepSeek R1 should be reserved for scenarios where its advanced reasoning justifies the resource investment. References Deepseek R1 model on Hugging Face https://huggingface.co/deepseek-ai/DeepSeek-R1 vLLM GitHub repository https://github.com/vllm-project/vllm Azure ND H100 v5 documentation https://learn.microsoft.com/en-us/azure/virtual-machines/nd-h100-v5-series FlashInfer GitHub repository https://github.com/flashinfer-ai/flashinfer DeepGEMM GitHub repository https://github.com/deepseek-ai/DeepGEMM AI-MO validation dataset on Hugging Face https://huggingface.co/datasets/AI-MO/aimo-validation-aime Appendix Install vLLM curl -LsSf https://astral.sh/uv/install.sh | sh uv venv myvllm --python 3.11 --seed source myvllm/bin/activate uv pip install vllm --torch-backend=auto git clone https://github.com/flashinfer-ai/flashinfer.git --recursive uv pip install ninja cd flashinfer uv pip install --no-build-isolation --verbose . Install DeepSeek DeepEP git clone https://github.com/vllm-project/vllm.git cd ~/vllm/tools/ep_kernels export CUDA_HOME=/usr/local/cuda-12.8 TORCH_CUDA_ARCH_LIST="9.0" (For Hopper) bash install_python_libraries.sh 2.&1 | tee install_python_libraries.log_$$ sudo bash configure_system_drivers.sh 2>&1 | tee configure_system_drivers.log_$$ sudo reboot Install DeepSeek DeepGEMM git clone --recursive https://github.com/deepseek-ai/DeepGEMM.git cd deepGEMM ./install.sh 2>&1 | tee install.log_$$ Configure DeepSeek R1 with vLLM on 2 ND_H100_v5 Second node configuration Execute this script on second node before the script on the primary node. #!/bin/bash MODEL="deepseek-ai/DeepSeek-R1" PORT=8000 export VLLM_LOGGING_LEVEL=INFO export HF_HUB_CACHE=/home/azureuser/cgshared/hf_cache #export VLLM_ALL2ALL_BACKEND=deepep_high_throughput export VLLM_ALL2ALL_BACKEND=deepep_low_latency export VLLM_USE_DEEP_GEMM=1 export GLOO_SOCKET_IFNAME=eth0 vllm serve $MODEL --port $PORT --tensor-parallel-size 1 --enable-expert-parallel --data-parallel-size 16 --data-parallel-size-local 8 --data-parallel-start-rank 8 --data-parallel-address 10.0.0.6 --data-parallel-rpc-port 23345 --headless --max-model-len 32768 --reasoning-parser deepseek_r1 Primary node configuration #!/bin/bash MODEL="deepseek-ai/DeepSeek-R1" PORT=8000 export VLLM_LOGGING_LEVEL=INFO export HF_HUB_CACHE=/home/azureuser/cgshared/hf_cache #export VLLM_ALL2ALL_BACKEND=deepep_high_throughput export VLLM_ALL2ALL_BACKEND=deepep_low_latency export VLLM_USE_DEEP_GEMM=1 export GLOO_SOCKET_IFNAME=eth0 vllm serve $MODEL --port $PORT --tensor-parallel-size 1 --enable-expert-parallel --data-parallel-size 16 --data-parallel-size-local 8 --data-parallel-address 10.0.0.6 --data-parallel-rpc-port 23345 --api-server-count 1 --max-model-len 32768 --reasoning-parser deepseek_r1 Install vLLM benchmark environment cd vllm uv pip install vllm[bench] Run vLLM benchmark #!/bin/bash vllm bench serve \ --backend vllm \ --model deepseek-ai/DeepSeek-R1 \ --endpoint /v1/completions \ --dataset-name hf \ --dataset-path AI-MO/aimo-validation-aime \ --ramp-up-strategy linear \ --ramp-up-start-rps 1 \ --ramp-up-end-rps 10 \ --num-prompts 400 \ --seed 42Inference performance of Llama 3.1 8B using vLLM across various GPUs and CPUs
Introduction Following our previous evaluation of Llama 3.1 8B inference performance on Azure’s ND-H100-v5 infrastructure using vLLM, this report broadens the scope to compare inference performance across a range of GPU and CPU platforms. Using the Hugging Face inference benchmarker, we assess not only throughput and latency but also the cost-efficiency of each configuration—an increasingly critical factor for enterprise deployment. As organizations seek scalable and budget-conscious solutions for deploying large language models (LLMs), understanding the trade-offs between compute-bound and memory-bound stages of inference becomes essential. Smaller models like Llama 3.1 8B offer a compelling balance between capability and resource demand, but the underlying hardware and software stack can dramatically influence both performance and operational cost. This report presents a comparative analysis of inference performance across multiple hardware platforms, factoring in: Token throughput and latency across chat, classification, and code generation workloads. Resource utilization, including KV cache utilization and efficiency. Cost per token, derived from cloud pricing models and hardware utilization metrics. By combining performance metrics with cost analysis, we aim to identify the most effective deployment strategies for enterprise-grade LLMs, whether optimizing for speed, scalability, or budget. Benchmark environment Inference benchmark The Hugging face Inference benchmarking code was used for the AI Inference benchmark. Three different popular AI inference profiles were examined. Chat: Probably the most common use case, question and answer format on a wide range of topics. Classification: Providing various documents and requesting a summary of its contents. Code generation: Providing code and requesting code generation, e.g. create a new function. Profile Data set Input prompt Output prompt Chat hlarcher/inference-benchmarker/share_gpt_turns.json N/A min=50, max=800, variance=100 Classification hlarcher/inference-benchmarker/classification.json Min=8000, max=12000, variance=5000 Min=30, max=80, variance=10 Code generation hlarcher/inference-benchmarker/github_code.json Min=3000, max=6000, variance=1000 Min=30, max=80, variance=10 Huggingface Lama 3.1 8B models used Precision Model Size (GiB) meta-llama/Llama-3.1-8B-Instruct FP16 14.9 vLLM parameters Default value gpu_memory_utilization 0.9 max_num_seqs 1024 max_num_batched_tokens 2048 (A100), 8192 (H100,H200) enable_chunked_prefill True enable_prefix_caching True VM Configuration GPU ND-H100-v5, ND-H200-v5, HD-A100-v4 (8 H100 80GB &40GB) running HPC Ubuntu 22.04 (Pytorch 2.7.0+cu128, GPU driver: 535.161.08 and NCCL 2.21.5-1). 1 GPU was used in benchmark tests. CPU Ubuntu 22.02 (HPC and Canonical/jammy) Results GPU Profile Avg prompt throughput Avg generation throughput Max # Requests waiting Max KV Cache usage % Avg KV Cache hit rate % H100 Chat ~2667 ~6067 0 ~14% ~75% Classification ~254149 ~1291 0 ~46% ~98% Code generation ~22269 ~266 ~111 ~93% ~1% H200 Chat ~3271 ~7464 0 ~2% ~77% Classification ~337301 ~1635 0 ~24% ~99% Code generation ~22726 ~274 ~57 ~46% ~1% A100 Chat ~1177 ~2622 0 ~2% ~75% Classification ~64526 ~333 0 ~45% ~97% Code generation ~7926 ~95 ~106 ~21% ~1% A100_40G Chat ~1069 ~2459 0 ~27% ~75% Classification ~7846 ~39 ~116 ~68% ~5% Code generation ~7836 ~94 ~123 ~66% ~1% Cost analysis Cost analysis used pay-as-you-go pricing for the south-central region and measured throughput in tokens per second to calculate the metric $/(1K tokens). CPU performance and takeaways The Huggingface AI-MO/aimo-validation-aime data was by vllm bench to test the performance of Llama 3.1 8B on various VM types (left graph below). It is a struggle (insufficient FLOPs and memory bandwidth) to run Llama 3.1 8B on CPU VM’s, even the best performing CPU VM (HB176-96_v4) throughput and latency is significantly slower than the A100_40GB GPU. Tips Enable/use AVX512 (avx512f, avx512_bf16, avx512_vnni etc) (See what is supported/available via lscpu) Put AI model on single socket (if it has sufficient memory). For larger models you can use tensor parallel to split the model across sockets. Use pinning to specify which cores the threads will run on (in vLLM, VLLM_CPU_OMP_THREADS_BIND=0-22) Specify large enough KVCache (on CPU memory). In vLLM, VLLM_CPU_KVCACHE_SPACE=100) Analysis Throughput & Latency H200 outperforms all other GPUs across all workloads, with the highest prompt and generation throughput. H100 is a close second, showing strong performance especially in classification and code generation. A100 and A100_40G lag significantly behind, particularly in classification tasks where throughput drops by an order of magnitude (on A100_40G, due to smaller GPU memory and lower KV Cache hit percentage). KV Cache Utilization H200 and H100 show efficient cache usage with high hit rates (up to 99%) and low waiting requests. (The exception is code generation which has low hit rates (~1%)) A100_40G suffers from high KV cache usage and low hit rates, especially in classification and code generation, indicating memory bottlenecks. The strain on the inference server is observed by the higher number of waiting requests. Cost Efficiency Chat profiles: The A100 GPU (40G) offers the best value. Classification profiles: The H200 is most cost-effective. Code-generation profiles: The H100 provides the greatest cost efficiency. CPU vs GPU Llama 3.1 3B can run on CPU VM’s but the throughput and latency are so poor compared to GPU’s if does not make an practical or financial sense to do so. Smaller AI models (<= 1B parameters) may be OK on CPU’s for some light weight inference serves (like Chat). Conclusion The benchmarking results clearly demonstrate that hardware choice significantly impacts the inference performance and cost-efficiency of Llama 3.1 8B deployments. The H200 GPU consistently delivers the highest throughput and cache efficiency across workloads, making it the top performer overall. H100 follows closely, especially excelling in code generation tasks. While A100 and A100_40G offer budget-friendly options for chat workloads, their limitations in memory and cache performance make them less suitable for more demanding tasks. CPU virtual machines do not offer adequate performance—in terms of throughput and latency—for running AI models comparable in size to Llama 3.1 8B. These insights provide a practical foundation for selecting optimal infrastructure based on inference workload type and cost constraints. References Hugging Face Inference Benchmarker https://github.com/huggingface/inference-benchmarker Datasets used for benchmarking: Chat: hlarcher/inference-benchmarker/share_gpt_turns.json Classification: hlarcher/inference-benchmarker/classification.json Code Generation: hlarcher/inference-benchmarker/github_code.json Model: meta-llama/Llama-3.1-8B-Instruct on Hugging Face https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct vLLM Inference Engine https://github.com/vllm-project/vllm Azure ND-Series GPU Infrastructure https://learn.microsoft.com/en-us/azure/virtual-machines/nd-series PyTorch 2.7.0 + CUDA 12.8 https://pytorch.org NVIDIA GPU Drivers and NCCL Driver: 535.161.08 NCCL: 2.21.5-1 https://developer.nvidia.com/nccl Azure Pricing Calculator (South-Central US Region) https://azure.microsoft.com/en-us/pricing/calculator CPU - vLLM Appendix Install vLLM on CPU VM’s git clone https://github.com/vllm-project/vllm.git vllm_source cd vllm_source edit Dockerfiles (vllm_source/docker/Dockerfile.cpu) cp Dockerfile.cpu Dockerfile_serve.cpu change last line to “ENTRYPOINT ["/opt/venv/bin/vllm","serve"]” cp Dockerfile.cpu Dockerfile_bench.cpu change last line to “ENTRYPOINT ["/opt/venv/bin/vllm","bench","serve"]” Build images (enable AVX512 supported features (see lscpu)) docker build -f docker/Dockerfile_serve.cpu --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_DISABLE_AVX512=false --tag vllm-serve-cpu-env --target vllm-openai . docker build -f docker/Dockerfile_bench.cpu --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_DISABLE_AVX512=false --tag vllm-bench-cpu-env --target vllm-openai . Start vllm server Remember to set <YOUR HF TOKEN> and <CPU CORE RANGE> docker run --rm --privileged=true --shm-size=8g -p 8000:8000 -e VLLM_CPU_KVCACHE_SPACE=<SIZE in GiB> -e VLLM_CPU_OMP_THREADS_BIND=<CPU CORE RANGE> -e HF_TOKEN=<YOUR HF TOKEN> -e LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD" vllm-serve-cpu-env meta-llama/Llama-3.1-8B-Instruct --port 8000 --dtype=bfloat16 Run vLLM benchmark Remember to set <YOUR HF TOKEN> docker run --rm --privileged=true --shm-size=4g -e HF_TOKEN=<YOUR HF TOKEN> -e LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD" vllm-bench-cpu-env --backend vllm --model=meta-llama/Llama-3.1-8B-Instruct --endpoint /v1/completions --dataset-name hf --dataset-path AI-MO/aimo-validation-aime --ramp-up-strategy linear --ramp-up-start-rps 1 --ramp-up-end-rps 2 --num-prompts 200 --seed 42 --host 10.0.0.4Performance of Llama 3.1 8B AI Inference using vLLM on ND-H100-v5
Introduction The pace of development in large language models (LLMs) has continued to accelerate as the global AI community races toward the goal of artificial general intelligence (AGI). Today’s most advanced models boast trillions of parameters, pushing the boundaries of what machines can understand and generate. However, this scale comes at a steep cost—both in terms of training and inference—due to the immense GPU resources required to host and operate these models. Yet, innovation is not limited to those with access to the largest AI supercomputers. DeepSeek have demonstrated that it is possible to build highly competitive models without relying on the latest, most expensive infrastructure. At the same time, a renewed wave of open-source collaboration is challenging the closed-source strategies of leading AI companies, offering more accessible and customizable alternatives. For enterprise customers, the focus is shifting toward practical, cost-effective solutions. Rather than deploying trillion-parameter giants, many organizations are turning to smaller models—such as those with around 8 billion parameters—that strike a balance between accuracy and efficiency. These models are not only easier to fine-tune and deploy but also significantly reduce the cost per token, making them ideal for real-world business applications. In this paper, we explore the capabilities of the Llama 3.1 8B model as a representative example of a modern, enterprise-ready LLM. We benchmark its inference performance on Azure’s ND-H100 v5 infrastructure using the vLLM engine and present our findings along with recommendations for enterprise deployment. AI Inference architecture Inference in transformer-based large language models (LLMs) is typically divided into two primary stages: prefill and decode. Understanding the characteristics and resource demands of each stage is essential for optimizing performance, especially when deploying models like Llama 3.1 8B in enterprise environments. Prefill Stage: Compute-Bound Initialization The prefill stage is responsible for processing the input prompt. It involves tokenizing the input and performing a forward pass through the model to populate the key-value (KV) cache. This stage is compute-intensive, as it requires full attention computation across all input tokens. The performance bottleneck here is typically the GPU's compute throughput, especially for long prompts or large batch sizes. Decode Stage: Memory-Bound Token Generation Once the KV cache is populated, the decode stage begins. This stage generates one token at a time, using the cached context to predict the next token. The decode step is memory-bound, as it relies heavily on fast access to the KV cache. When the model achieves a KV cache hit, it can skip re-computation of earlier tokens, significantly reducing latency and compute load. This makes cache efficiency a critical factor in real-time inference performance. Fig 1. High level architecture of AI Inference, showing efficient use of KV cache can increase token throughput and reduce AI inference latency. Overall Inference Characteristics In general, AI inference is memory-bound, particularly during the decode phase. The ability to serve multiple concurrent requests efficiently depends on how well the system can manage memory bandwidth and cache locality. As such, optimizing memory access patterns and minimizing cache misses are key to achieving high throughput and low latency. Techniques for Optimization To maximize GPU utilization and token throughput while minimizing latency, several architectural strategies are employed: KV Cache Management: Efficient reuse and eviction policies to maintain high cache hit rates. vLLM uses PagedAttention (which is inspired by virtual memory and paging techniques used in operating systems), to manage the KV cache using blocks/pages. This allows vLLM to efficiently/dynamically utilize HBM memory and minimizes memory fragmentation. Batching and Scheduling: Grouping similar requests to improve parallelism and reduce idle GPU time. VLLM has a few parameters to control batching/parallelism. MAX_NUM_SEQS: How many input requests to process in parallel. MAX_NUM_BATCHED_TOKENS: The number of tokens to process in parallel (Forward pass in Deep neural network) Note: Larger values may not always be optimal; you could improve token throughput at the expense of latency. Weight and Activation Quantization (fp8): Reducing the precision of AI model weights and activations can give more memory to load AI models or have a larger KV cache. Lowering the precision also allows computations to be performed on more efficient GPU (higher FLOPS) computational units. Parallelization techniques: Tensor parallelism, Pipeline parallelism, Expert parallelism or Data parallelism can be used to split larger models across multiple Nodes/GPUs. Tensor parallelism distributes the model across the GPUs, with each GPU handling multiple layers of the model. Pipeline parallelism involves dividing the model, where a group of nodes (or GPUs) is responsible for processing its assigned DDN layer. Expert parallelism supports Mixture of Experts (MoE) models where different expert networks can be distributed across GPU’s. Data parallelism replicates the entire model across multiple GPU sets and processes different batches of requests in parallel Speculative Decoding: Predicting multiple tokens ahead to reduce the number of forward passes. Prefill/Decode Decoupling: Recent advancements, such as those implemented in vLLM (and NVIDIA Dynamo), decouple the prefill and decode stages, allowing each to be assigned dedicated GPU or CPU resources. This separation enables better resource allocation and parallelism, especially in multi-tenant or high-throughput environments. By leveraging these techniques, vLLM provides a highly efficient inference engine that is well-suited for serving modern LLMs like Llama 3.1 8B. This makes it a compelling choice for enterprise applications where cost, latency, and scalability are critical considerations. Benchmark environment Inference benchmark The Huggingface Inference benchmarking code was used for the AI Inference benchmark. Three different popular AI inference profiles were examined. Chat: Probably the most common use case, question and answer format on a wide range of topics. Classification: Providing various documents and requesting a summary of its contents. Code generation: Providing code and requesting code generation, e.g. create a new function. Profile Data set Input prompt Output prompt Chat hlarcher/inference-benchmarker/share_gpt_turns.json N/A min=50, max=800, variance=100 Classification hlarcher/inference-benchmarker/classification.json Min=8000, max=12000, variance=5000 Min=30, max=80, variance=10 Code generation hlarcher/inference-benchmarker/github_code.json Min=3000, max=6000, variance=1000 Min=30, max=80, variance=10 Huggingface Lama 3.1 8B models used Precision Model Size (GiB) meta-llama/Llama-3.1-8B-Instruct FP16 14.9 neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8 FP8 8.4 nvidia/Llama-3.1-8B-Instruct-FP8 FP8 8.4 vLLM parameters Default value gpu_memory_utilization 0.9 max_num_seqs 1024 max_num_batched_tokens 8192 enable_chunked_prefill True enable_prefix_caching True Dynamo parameters Default values block-size 64 max-model-len 16384 kv_connector DynamoNixlConnector router round-robin remote-prefill True conditional-disagg True max-local-prefill-length 10 max-prefill-queue-size 2 max-num-batched-tokens 16384 Environment Local No-operation (Planner) True Results Fig 2: AI Inference performance comparison between the chat, classification and code generation profiles on ND-H100-v5 (8 H100). Profile Avg prompt throughput Avg generation throughput Max # Requests waiting Max KV Cache usage % Avg KV Cache hit rate % Chat ~7500 ~17600 0 ~2% ~78% Classification ~55300 ~2900 0 ~5% ~100% Code generation ~130400 ~1450 ~ 37 ~10% ~2% Fig 3: Show the impact of modifying MAX_NUM_BATCHED_TOKENS on the code-generation inference benchmark. (This parameter has a greater impact on the code generation benchmark compared to the chat/classification because of the low KV cache hit percentage.) Fig 4: Code generation inference benchmark run on 1 H100 showing the performance impact of using fp8 quantization. Fig 5: Code generation benchmark run on 1, 2, 4 & 8 H100. The results indicate that higher token throughput could be achieved running inference using 1 copy of model on each GPU instead of distributing model (via tensor parallel) amongst the 8 GPUs. Fig 6: Impact on throughput (tokens/sec) by adjusting AI Inference configuration (vLLM) on ND-H100-v5. Results comparing Dynamo vs traditional vLLM Fig 7: Dynamo vs traditional vLLM throughput (tokens/sec) comparison on 1 ND_H100_v5 (8 GPU’s). The best traditional vLLM configuration (8 x vLLM tensor_parallel=1) throughput performance is compared with various Dynamo configurations (i.e different combinations of GPU’s assigned for decode and prefill). Note: vLLM does have an experimental disaggregated prefill capability with various connector types. I attempted to run vLLM experimental disaggregated prefill using the kv-connector = LMCacheConnectorV1 (with Nixl enabled). I got mixed results, eventually running into the following issues (and deciding to switch to Nvidia Dynamo instead). Limited control over allocating GPU’s to decode vs prefill (used tensor parallel option, but limited by specific ratio with number of attention multi-heads). Memory management problems, got OOM errors even though there was plenty of HBM memory available on GPU’s (HBM was not distributed evenly amongst GPU’s). Analysis The performance of the inference prefill is determined by length and number of input prompts, this phase is compute bound (effectively a Matrix-Matrix operation) and we see that the code-generation profile does best primarily because it had the largest number of input tokens. The decode phase is a memory bound operation (effectively a Vector-Matrix operation) and performance in this phase is heavily dependent on the KV cache hit percentage. The code-generation profile only had ~1.7% KV cache hit percentage (there was plenty of HBM capacity, only ~10% of available KV Cache capacity was used), which resulted in slow decode performance impacting its overall throughput and latency especially at higher QPS (the code-generation benchmark was the only one which had requests being backed-up and waiting). The classification profile did well in the decode phase primarily due to the high KV cache hit percentage (~100%), it did struggle in overall throughput due to the small length of the output tokens. Adjusting the size of MAX_NUM_BATCHED_TOKENS had very little impact on the chat and classification benchmarks probably because they had high KV Cache hit percentages, but it did impact the performance of the code-generation benchmark (a ~3% improvement in tokens/sec using a larger value). Quantization of AI model can free up HBM memory to allow a large model to load or improve the AI inference performance by providing more memory for KV caching, it can also improve performance by performing computations with a higher FLOPS lower precision (e.g. FP8) format. In this case there is plenty of HBM and all three profiles did not use all the available KV cache space available, so FP8 quantization does not improve KV caching efficiency. Improvements in compute performance are observed with quantization, especially for the code generation profile which had a low KV cache hit percentage. The code generation tokens/sec on 1 GPU improved by ~38%. Since the Llama 3.1 8B model can easily fit on 1 H100, you can get significantly better total throughput (tokens/sec) on ND-H100-v5 if a complete model is loaded into each separate GPU instead of splitting the model across multiple GPU’s (via tensor parallel). The chat, classification and code generation inference throughput improved 4.2, 5.2, 1.9 times respectively. Newer inferencing server architectures feature disaggregated prefill, which allows you to decouple prefill from decode and assign resources (GPU’s, CPU’s) to each type of worker (prefill or decode). This is especially suited for large reasoning models with large context windows, running on large GPU inference clusters, significant performance gains have been reported. In this case we have a modest size (8B parameters) NLP model running on a single ND_H100_v5 node (only 8 GPU’s), so we were not expecting any significant performance improvements. The traditional aggregated vLLM was much faster than the best Dynamo configuration, running this inference benchmark on llama 3 8B model on ND_H100_v5. In this case the model can fit in a single GPU and the overhead of disaggregation might outweigh any parallelism gains when one GPU can already handle both phases efficiently. Conclusions When analyzing the performance of AI inference, it’s important not only to focus on the number of input and output tokens but also on the type of AI inference application, which can have a big impact on KV Cache effectiveness. Smaller AI models provide opportunities and more options to configure your environment in an optimal way to maximize token/sec throughput. References Welcome to vLLM — vLLM Mastering LLM Techniques: Inference Optimization | NVIDIA Technical Blog huggingface/inference-benchmarker: Inference server benchmarking tool meta-llama/llama-models: Utilities intended for use with Llama models. [2309.06180] Efficient Memory Management for Large Language Model Serving with PagedAttention ND-H100-v5 size series - Azure Virtual Machines | Microsoft Learn Hugging Face – The AI community building the future. Introducing NVIDIA Dynamo, A Low-Latency Distributed Inference Framework for Scaling Reasoning AI Models | NVIDIA Technical Blog Appendix Installation of hugging face inference benchmarker Install Rust curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh Build and install inference benchmarker cargo install --git https://github.com/huggingface/inference-benchmarker/ Installation of vLLM Install uv curl -LsSf https://astral.sh/uv/install.sh | sh Set-up python workspace uv venv --python 3.10 --seed source .venv/bin/activate Install vLLM uv pip install vllm --torch-backend=auto Install FlashInfer git clone https://github.com/flashinfer-ai/flashinfer.git --recursive pip install ninja cd flashinfer pip install --no-build-isolation --verbose . Start vLLM server #!/bin/bash NUM_GPUS=8 #NUM_GPUS=4 #NUM_GPUS=2 #NUM_GPUS=1 MODEL=meta-llama/Llama-3.1-8B-Instruct #MODEL=nvidia/Llama-3.1-8B-Instruct-FP8 #MODEL=neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8 PORT=5000 export VLLM_LOGGING_LEVEL=INFO python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port $PORT --model $MODEL --tensor-parallel-size $NUM_GPUS --dtype auto Run Inference benchmark TOKENIZER_NAME="meta-llama/Llama-3.1-8B-Instruct" #TOKENIZER_NAME="neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8" #TOKENIZER_NAME="nvidia/Llama-3.1-8B-Instruct-FP8" inference-benchmarker \ --tokenizer-name $TOKENIZER_NAME \ --url http://localhost:5000 \ --profile code-generation -n Nvidia Dynamo installation Install python virtual environments sudo apt install python3.10-venv Create dynamo virtual workspace python3 -m venv mydynamo Activate virtual environment source /home/azureuser/mydynamo/bin/activate Check out dynamo github code. git clone https://github.com/ai-dynamo/dynamo.git cd dynamo git checkout $(git describe --tags $(git rev-list --tags --max-count=1)) Install dynamo prerequisites pip install "ai-dynamo[all]" Install some additional python modules and packages pip install tensorboardX pip install pynvml sudo apt install etcd systemctl status nats Restart etcd systemctl restart etcd Set-up nats-server wget https://github.com/nats-io/nats-server/releases/download/v2.10.22/nats-server-v2.10.22-linux-amd64.zip unzip nats-server-v2.10.22-linux-amd64.zip sudo mv nats-server-v2.10.22-linux-amd64/nats-server /usr/local/bin/ sudo chmod +x /usr/local/bin/nats-server Create/edit /etc/systemd/system/nats.service [Unit] Description=NATS Server After=network.target [Service] Type=simple ExecStart=/usr/local/bin/nats-server -js Restart=always RestartSec=10s LimitNOFILE=40000 [Install] WantedBy=multi-user.target export NATS_SERVER="nats://localhost:4222" Start Dynamo server/service cd $DYNAMO_HOME/examples/llm edit disagg.yaml file to modify parameters. dynamo serve graphs.disagg:Frontend -f ./configs/disagg.yaml Set Dynamo environmental variables export DYNAMO_HOME=$(pwd) cd $DYNAMO_HOME/examples/llmOptimizing Large-Scale AI Performance with Pretraining Validation on a Single Azure ND GB200 v6
Small performance gaps on a single virtual machine lead to large and costly performance losses at scale. Running small-scale pretraining jobs enables single-VM validation and allows for fine-grained identification of issues such as performance degradation, hardware bottlenecks, or software inefficiencies ahead of large-scale runs. In this post, we present a practical methodology for benchmarking the ND GB200 v6 virtual machines (VMs). A single ND GB200 v6 VM on Azure is powered by two NVIDIA Grace CPUs and four NVIDIA Blackwell GPUs. To ensure reliability in production workloads we used automated pretraining of lightweight Llama models with the NVIDIA NeMo framework. By systematically exploring and tuning key performance parameters and rigorously cross-checking results with detailed telemetry, we identify conditions that most significantly stress the GPUs. You can reproduce and reuse these pretraining workloads from our fully automated Azure AI Benchmarking guide.
