Best HPC Software

OpenPBS is an open source workload manager and job scheduler for high performance computing environments, covering clusters, clouds and supercomputers, with the stated aim of improving both system efficiency and people's productivity. Those two goals pull in different directions in practice, since a scheduler tuned purely for machine utilisation can leave researchers waiting, and the project's framing acknowledges both sides.

Its origins matter to how it is regarded. The project describes itself as built by HPC people for HPC people, and the PBS lineage is long established in scientific and academic computing, which means job scripts, site policies and operator expertise already exist for it across a great many institutions. For a site choosing a scheduler, the availability of people who already know it is a substantial practical consideration alongside features.

As open source software there is no licence cost, and the project publishes everything needed to adopt it independently: downloads, source code, an issue tracker, a community forum and a contributors portal. Commercial support is available separately for organisations that need a support contract behind production infrastructure, which is the arrangement most institutional deployments require regardless of licensing.

Note the ownership context. Altair, which stewards OpenPBS, has been acquired by Siemens, with the commercial PBS Professional product now presented as part of the Siemens HPCWorks portfolio. OpenPBS itself remains the open source project, but anyone planning a multi-year deployment should watch how the open source and commercial editions are positioned going forward. Documentation is published in English and Japanese.

Google Cloud Cluster Toolkit deploys high performance computing clusters on Google Cloud using a blueprint model, and that model is the substance of it. Rather than configuring infrastructure directly, a cluster is described in a blueprint which is then turned into deployment files and provisioned. A published cluster blueprint catalogue provides ready-made starting points, so a working cluster comes from adapting a known-good definition rather than assembling one from primitives.

The reproducibility that follows matters for research computing specifically. A blueprint held in version control recreates the same environment later, which supports reproducible results in a way that a hand-built cluster cannot. Modules compose blueprints from reusable pieces, so an organisation can standardise on its own building blocks across multiple clusters.

The published quickstarts show where Google is directing this, and it is as much AI as traditional HPC. Named paths include deploying an HPC cluster with Slurm, creating an AI-optimised Slurm cluster with an A4 machine type, and creating an RDMA-enabled HPC Slurm cluster with H4D. RDMA support is significant for genuine HPC work, since tightly coupled parallel jobs are limited by interconnect latency rather than raw compute, and it is often what separates a usable cluster from an expensive one.

Slurm appears throughout as the scheduler, which is consistent with the rest of this category: teams bring existing Slurm expertise rather than learning something new. Documentation covers product overview, blueprints, deployment files, modules, quickstarts, dependency installation and workload deployment. The toolkit can be started free, with cost following Google Cloud consumption.

Azure CycleCloud is an enterprise-friendly tool for orchestrating and managing high performance computing environments on Azure. It provisions the infrastructure for HPC systems, deploys familiar HPC schedulers onto it, and automatically scales that infrastructure so jobs run efficiently at any scale. The word familiar is doing real work in that description: HPC teams have years of accumulated job scripts, submission habits and operational knowledge tied to specific schedulers, and a cloud tool that requires abandoning all of it faces resistance no feature list overcomes.

The orchestration model separates two concerns that are usually tangled together. Infrastructure provisioning handles the compute, storage and networking Azure needs to stand up. Scheduler deployment then installs and configures the workload manager on top. Keeping these distinct is what allows the same team to run different schedulers for different workloads, or to migrate an on-premises cluster's scheduler configuration into the cloud without redesigning it.

Automatic scaling is the economic argument for cloud HPC generally. An on-premises cluster is sized for peak demand and idle much of the time, whereas infrastructure that scales with the job queue is paid for only while working. Whether that saves money depends entirely on utilisation patterns, so the honest evaluation is to compare your actual queue depth over time against fixed hardware cost rather than assuming cloud is cheaper.

Documentation is published through Microsoft Learn with full guides. Note that access to some documentation pages requires authorisation. Pricing follows Azure consumption rather than a separate licence for the tool.

AWS ParallelCluster is an open source cluster management tool for deploying and managing high performance computing clusters on AWS. Its commercial position is unusually simple: the tool is available at no additional charge, and you pay only for the AWS resources needed to run your applications. For a research group or engineering team, that removes the licence question entirely and leaves only the compute bill, which is the opposite of the traditional HPC scheduler model where licensing is a substantial line item.

Clusters are modelled and provisioned either through a simple graphical interface or a text file, in an automated and secure manner. The text file route is what makes clusters reproducible: a configuration in version control can recreate an identical environment, which matters for research reproducibility as much as for operations. Automatic resource scaling then provisions and dynamically scales resources according to the workload rather than holding a fixed cluster.

Scheduler choice is a genuine strength. ParallelCluster supports Slurm, the dominant open source HPC scheduler, as well as AWS Batch, so teams with existing Slurm job scripts and expertise can move to cloud infrastructure without rewriting how work is submitted. Multiple instance types and job submission queues are supported, allowing different classes of work to be routed to appropriate hardware.

The project is genuinely open: built on the CfnCluster project, distributed through an AWS CloudFormation template or the Python Package Index, with source code hosted on the AWS GitHub repository. AWS also offers Parallel Computing Service as a managed Slurm alternative for teams that would rather not operate the cluster themselves.

Vantage Compute is a control plane for running compute workloads across environments, covering AI, HPC and quantum work. Its central technical proposition is bringing Slurm to Kubernetes, so existing HPC workloads run without refactoring or compromise. That addresses a genuine and awkward split in current infrastructure: organisations have standardised operations on Kubernetes, while their HPC workloads are written for Slurm, and reconciling the two usually means running two separate stacks with separate teams.

Cluster provisioning is fully automated, described as being up and running with a few clicks rather than through a build project. For teams whose bottleneck is the weeks required to stand up a cluster rather than the compute itself, that is the relevant improvement.

Identity-aware infrastructure is the other distinguishing element and is more consequential than it sounds. The platform provides secure identity-based connectivity across Slurm, AI tools and Kubernetes, with identity-based authentication eliminating passwords across compute, storage and workloads. Credential sprawl across HPC estates is a persistent security weakness, since shared accounts and long-lived keys are common and rarely rotated, so replacing them with identity-based access closes a real gap rather than adding a feature.

The platform is described as GPU-native, reflecting that AI workloads rather than traditional simulation now drive much HPC investment. Access is available through a native user interface, a command line interface and an API, so both interactive users and automated pipelines are served. Pricing is published on the vendor's site, and evaluation begins with a booked demo. Documentation is available openly.

IBM Spectrum LSF Suites is a workload management platform and job scheduler for distributed high performance computing. It sits alongside PBS Professional and Slurm as one of the long-established schedulers in this field, with the institutional familiarity and accumulated site expertise that implies.

The most useful thing for anyone evaluating it is that a Community Edition is available to download free. That allows the scheduler to be assessed properly against real jobs before any commercial conversation, which is unusual for enterprise HPC software and puts it on comparable footing with the open source alternatives for evaluation purposes.

Cloud provisioning is addressed concretely rather than aspirationally. Terraform-based automation is available to provision and configure the resources for an LSF-based cluster on IBM Cloud, which means a cluster is defined as code and reproducible rather than assembled by hand. Dynamic hybrid cloud support allows workloads to be forwarded to multiple environments, so on-premises capacity can be supplemented rather than replaced, which is how most established HPC sites actually adopt cloud.

The user experience receives more attention than is typical for HPC schedulers, where the interface is often an afterthought behind command-line submission. A fully integrated UI is offered with mobile clients for job monitoring and notifications and an integrated desktop client for Windows environments. Windows support matters in engineering and simulation settings where the analysts submitting work are not Linux users. Sites can add their own customisations and extensions. A published roadmap is available. Pricing is not published for the commercial suites.

PBS Professional is a commercial workload manager and job scheduler for high performance computing, and its branding has changed: Altair has been acquired by Siemens, and the product is now presented as HPCWorks PBS Professional within the Siemens portfolio. Anyone searching for the Altair name will find it there, and older documentation and reviews refer to the previous branding.

The product addresses three outcomes stated plainly: improving productivity, optimising utilisation and simplifying administration, across clusters, clouds and supercomputers. Its stated range is unusually wide, spanning the largest high performance computing workloads down to millions of small, high-throughput jobs. That span matters because scheduling a handful of enormous parallel jobs and scheduling millions of tiny ones are genuinely different problems, and many schedulers handle one well and the other poorly.

Policy-driven scheduling is the core mechanism, used to optimise resources and prioritise critical jobs. Dynamic policies allow the rules to change with circumstances rather than being fixed at configuration time, which is what a shared institutional cluster needs when priorities shift between departments or projects. A powerful customisation API supports site-specific extensions, and extensive hardware support covers heterogeneous estates.

Mixed AI and HPC workloads are called out specifically, reflecting the current reality that GPU-heavy AI training now shares clusters with traditional simulation work, and the two have different scheduling characteristics. The vendor's supporting claim is proven software backed by decades of experience across a broad range of industries with world-class support, which is the practical argument for a commercial scheduler over the free alternatives. Pricing is not published.

HPCWorks is a portfolio rather than a single product, covering high performance computing on premises and in the cloud. It is the umbrella under which PBS Professional and related tools now sit, and its branding has changed: Altair has been acquired by Siemens, and HPCWorks is presented within the Siemens Xcelerator platform as part of what Siemens describes as the most complete AI-powered portfolio of industrial software for simulation, high performance computing, data science and artificial intelligence.

The stated purpose is managing IT complexity, streamlining administration, controlling costs and enabling AI and mixed workloads. Cost control deserves attention as a distinct goal: HPC infrastructure is expensive whether owned or rented, and the question of which jobs justified which spend is one most sites answer poorly. A portfolio that treats cost visibility as a first-class concern rather than a reporting afterthought is addressing a real gap.

The application range cited is deliberately broad, spanning healthcare, weather prediction and chip design, which are three of the most computationally demanding fields and have quite different job profiles. GPU acceleration and rapid scaling support current AI workloads, with flexible scheduling and workflow design.

Agentic AI appears in an unusual and practical application: estimating job requirements. Users are notoriously bad at predicting how long a job will run or how much memory it needs, and inaccurate requests are a primary cause of poor cluster utilisation, since over-requested resources sit idle and under-requested jobs fail. Using AI to improve those estimates addresses utilisation at its actual source rather than through scheduler tuning alone. Pricing is not published and contact runs through the vendor.