What is HPC (high-performance computing)?

HPC systems are assembled rather than bought as a unit:

• Compute nodes - machines with many cores and large memory footprints, frequently identical so jobs scale predictably across them.
• Interconnect - the fast fabric between nodes, on which tightly coupled jobs depend more than on anything else.
• Storage - parallel filesystems and scratch space sized to keep data near the nodes consuming it.
• A batch scheduler - the queue that admits jobs, allocates nodes and decides what runs next.
• Site services - power, cooling and physical access, which are operations rather than software but decide uptime all the same.

The hardware is the straightforward half. A cluster earns its keep when jobs arrive in a shape the scheduler can place: data staged close to compute, submissions written for the queue, failures expected and recovered from checkpoints rather than from the beginning.

Two of those are habits rather than features. Staging data near compute and writing submissions for the queue are choices the submitting team makes, and they separate a cluster that stays busy from one that spends its day waiting on jobs that could have started hours earlier. Hardware does not compensate for either mistake; it simply makes the wait more expensive.

Tightly coupled jobs need whole nodes and a fast fabric, because every participant waits on the others; throughput jobs simply fill cores and care little about latency between nodes. Both are admitted through a batch queue rather than by requesting machines one at a time, and SLURM is the scheduler most often found in that role - submit, wait for allocation, run, release.

That queueing pattern is why HPC procurement has historically been a capacity decision rather than an on-demand one: a site buys for its peak workload and the scheduler smooths demand across it. The operational side of running this on VirtEngine - schedulers, adapters and job-shaped orders - is set out in HPC on VirtEngine.

The distinction shapes how capacity is planned. Throughput work spreads across whatever machines happen to be free and tolerates a busy queue; tightly coupled work cannot, because a single slow participant stalls everyone waiting on it. Sites respond by reserving the fastest fabric for jobs that will genuinely use it and pushing everything else to the general pool. Renters of the same systems face an echo of the question: whether a job needs a whole machine or only a share of one.

Universities and research facilities simulate weather, materials, fluids and molecular behaviour; industry teams model vehicles, structures and risk; media studios render; and machine-learning training has become a frequent guest on the same machines, since it also wants many accelerators and a scheduler.

What these buyers share is a queue-shaped demand: occasional need for a lot of compute, deadlines that vary, and no desire to own a cluster that would sit between runs. That pattern is what makes per-job purchasing worth considering at all.

The other shared trait is unevenness. Work arrives in bursts around deadlines with quiet stretches in between, and a cluster owned outright is paid for during those stretches as well. A large site absorbs that by keeping one queue full across many groups; a single team cannot, which is why shared facilities and rented capacity exist as answers to the same arithmetic.

On VirtEngine, an HPC requirement is written down rather than assumed: node count, duration, interconnect and storage expectations, plus the scheduler the job is written for. The tenant funds an escrow account with that request, so providers with free nodes can see committed budget before answering it, and bids compete on price and on what the site can actually offer.

Once matched, the lease is fulfilled through the provider's scheduler adapter, which places the job on real hardware and reports node-time as signed usage records. Settlement draws validated line items from escrow after the dispute window and returns whatever is unspent. Capacity and category detail sit with HPC listings on the market.