How Much Does AWS ParallelCluster Cost to Run? Real Numbers

TL;DR
- AWS ParallelCluster has no license or management fee. Everything you pay for is the EC2, EBS, FSx for Lustre, NAT gateway and log traffic the cluster creates — compute instances usually land at 85–95% of the total.
- The head node runs 24/7 and is not part of the Slurm fleet, so it bills even when zero jobs run. At m5.xlarge on-demand in us-east-1 that floor is roughly $140/month before you compute anything.
- Scale is what moves the number, not the cluster software. The same 20-node queue costs roughly $5–7k/month on Spot and roughly $14k/month on-demand at 8 job-hours × 22 days — a 2.5x swing from one config flag.
What the cluster bills you for (and what it doesn't)
ParallelCluster is an orchestrator: CloudFormation templates, a Slurm config, and a pcluster CLI that wires them together. AWS doesn't meter it. I've built clusters for genomics and CFD teams and never seen a line item called "ParallelCluster" on a bill — the surprise always comes from something the template spawned.
The billable pieces are predictable once you know the shape:
| Line item | Billing unit | Order of magnitude, us-east-1, 2026-11 |
|---|---|---|
| Head node (m5.xlarge) | $/hr, 730 hr/mo | ~$140/mo on-demand, 24/7 |
| Compute (c5.xlarge) | $/hr per node | ~$0.17/hr on-demand; Spot often 60–70% less |
| Compute (c5n.18xlarge) | $/hr per node | ~$3.9/hr on-demand |
| EBS gp3 root volumes | $/GB-month | ~$0.08/GB-month |
| FSx for Lustre (SSD scratch) | $/GB-month provisioned | ~$0.14/GB-month |
| NAT gateway | $0.045/hr + $0.045/GB | ~$33/mo just for existing |
| CloudWatch Logs | $/GB ingested | cheap until Slurm gets chatty |
| ParallelCluster itself | — | $0 |
Verify every rate on the current pricing pages before you quote a client — us-east-1 vs eu-central-1 vs ap-southeast-2 differ enough to break a budget, and Savings Plans change the EC2 column significantly.
Three cluster shapes, three different orders of magnitude
Estimates below assume a single queue, gp3 root volumes, and 22 working days. They're formulas with numbers plugged in, not quotes.
- SSL域名证书申请 · 小程序证书 · 国外SSL证书
- 代购 SSL 证书 · DigiCert / GeoTrust / RapidSSL 单域名与通配符
- SSL证书申请 · SSL证书购买 · SSL证书安装配置
- 服务器维护 · 网站代维护 · 服务器托管代维
- SSL证书安装服务 · 腾讯云 HTTPS 部署 · 证书过期更新
- NGINX / APACHE / TOMCAT / IIS 证书配置 · SSL证书安装
| Setup | Head node | Compute fleet | Duty cycle | Monthly order of magnitude |
|---|---|---|---|---|
| A — dev/course cluster | 1× m5.xlarge | 0–8× c5.xlarge Spot | 4 h/day | $150–$300 |
| B — CAE batch shop, Spot | 1× m5.2xlarge | 20× c5n.18xlarge Spot | 8 h/day | $5,000–$7,000 |
| C — same shop, on-demand | 1× m5.2xlarge | 20× c5n.18xlarge on-demand | 8 h/day | $13,500–$15,000 |
Setup A is the one most people start with: a head node that idles at $140/month and a fleet that only exists during class hours. Setup B adds a 1.2 TB FSx for Lustre scratch filesystem (~$170/month) and 1 TB of gp3 (~$80/month), which is noise next to the compute line. Setup C is the same cluster built by someone who left CapacityType: ON_DEMAND in the template.
For an MPI-shaped workload, swap in hpc6a.48xlarge at roughly $2.88/hr on-demand: 32 nodes × 6 h × 20 days is 3,840 node-hours, so somewhere near $11k/month on-demand and a third of that on Spot.
The formula I run before every build
I keep this in a shell script and paste new numbers in. It takes ten seconds and has stopped three bad estimates from reaching a client deck.
# monthly_estimate.sh — order of magnitude only, not a quote
HEAD_RATE=0.192 # m5.xlarge on-demand us-east-1, verify current price
HEAD_HOURS=730
NODES=20
HOURS_PER_DAY=8
DAYS=22
COMPUTE_RATE=1.45 # c5n.18xlarge Spot, check the Spot price history first
STORAGE_GB=1200
STORAGE_RATE=0.14 # FSx for Lustre SSD scratch
head=$(echo "$HEAD_RATE*$HEAD_HOURS" | bc)
compute=$(echo "$NODES*$HOURS_PER_DAY*$DAYS*$COMPUTE_RATE" | bc)
storage=$(echo "$STORAGE_GB*$STORAGE_RATE" | bc)
echo "head=$head compute=$compute storage=$storage total=$(echo "$head+$compute+$storage" | bc)"
The inputs that matter are node count, wall-clock hours per job day, and the rate for the instance family you actually need. Everything else is rounding error. Note that "hours per job day" is wall-clock, not CPU-hours — a 6-hour CFD run that uses 32 cores is 6 node-hours on one c5n.18xlarge, not 192.
Two more multipliers people forget: inter-node traffic crossing an Availability Zone (billed at $0.02/GB each direction in us-east-1) and EBS snapshots if you keep AMIs around. A cluster spread across two AZs for capacity reasons can quietly add hundreds per month on a data-heavy MPI job.
## Sensitivity factors that swing the bill 10x
**Spot vs on-demand** is the biggest lever by an order of magnitude, but it's also where I've seen the job that never finishes. Spot interruption on a tightly coupled MPI job means restarting from the last checkpoint. If your code doesn't checkpoint every 15–30 minutes, the retry time can cost more than the discount saves.
**`MinCount` on a compute queue.** This is the one that bit us. Setting `MinCount: 8` tells Slurm to keep eight nodes running permanently — it prevents scale-to-zero in a way that's easy to forget after a deadline sprint.
**`ScaledownIdletime`** (default 10 minutes) controls how long an idle node burns before Slurm drains it. With a fleet of 20 large nodes, 10 minutes of idle tail after every job wave adds up to real hours over a week.
**FSx for Lustre type.** Scratch is roughly half the per-GB rate of persistent SSD, and you should not be paying for persistent on a cluster you tear down on Fridays.
**Capacity, not price, can be the actual constraint.** Twice this year I've seen a queue sit at zero nodes while jobs waited:
An error occurred (InsufficientInstanceCapacity) when calling the RunInstances
operation: We currently do not have sufficient c5n.18xlarge capacity in the
Availability Zone you requested (us-east-1a).
The cost impact is indirect but real — idle researchers, a head node billing while nothing runs, and someone eventually hard-coding a bigger on-demand instance type to make it stop.
## There is no free tier for real clusters
You can create a tiny cluster that touches no free-tier allowance worth having. The defaults already break it: root volumes are sized in the tens of GB (gp3) and the 30 GB free-tier EBS allowance doesn't cover them, NAT gateways never have a free tier, and FSx for Lustre has no free usage at all.
The honest free-tier answer: you can run `pcluster configure` and inspect the generated YAML for free, and you can validate a template with `pcluster validate-config`. The first `pcluster create-cluster` that actually succeeds will cost you money within the hour.
## When NOT to use ParallelCluster
If your job fits on one instance — say a 32-vCPU c5.9xlarge for 40 minutes — a single EC2 instance or a Batch job is cheaper and simpler, because you skip the $140/month head node entirely. ParallelCluster starts paying off when you have a queue of jobs competing for many nodes and need a shared filesystem across them.
Skip it too if nobody on the team will maintain Slurm, or if the workload is bursty enough that the cluster would be stopped more than it runs. In that case an on-demand fleet managed by AWS Batch, or a managed HPC service, avoids the always-on head node.
## Hard-won lesson: the long weekend that cost $2,200
2026-11-20, Friday afternoon. We built a genomics cluster with a compute queue pinned at `MinCount: 8` because the customer wanted "nodes ready the moment the pipeline fires". The assembly job finished Saturday morning. Nobody stopped anything.
Monday morning Cost Explorer showed about **$2,190** for that cluster across three days: 8 nodes × 72 hours × ~$3.89/hr on-demand, plus the head node. The compute fleet had been running at zero jobs since Saturday.
My first assumption was wrong — I blamed the head node and checked its instance type. The head node was an m5.xlarge doing nothing unusual. The real state was in the fleet:
```bash
$ pcluster describe-compute-fleet --cluster-name genome-prod --region us-east-1
{
"Status": "RUNNING",
"LastStatusUpdatedTime": "2026-11-21T09:14:02.000Z"
}
$ sinfo
PARTITION AVAIL TIMELIMIT NODES STATE NODELIST
compute* up infinite 8 idle compute-dy-c5n18xlarge-[1-8]
`idle` with a real node list means running and billing; `idle~` means powered down. Eight `idle` nodes at noon on a Monday is the signature of a non-zero `MinCount`. Verification after the fix: we set `MinCount: 0`, redeployed, and watched `sinfo` return `idle~` within one `ScaledownIdletime` window. Cost Explorer's daily line for that cluster dropped to the head node plus the few hours jobs actually ran.
The lesson wasn't "use Spot" — it was that `MinCount` is a standing commitment, and it belongs in the same review as the budget alarm.
## Checks that keep the bill honest
Run these before every weekend and every demo.
```bash
# 1. Stop the compute fleet when you're done — head node stays up
pcluster update-compute-fleet
--cluster-name hpc-prod
--region us-east-1
--status STOPPED
# 2. Confirm the fleet state and node count
pcluster describe-compute-fleet --cluster-name hpc-prod --region us-east-1
pcluster list-clusters --region us-east-1
# 3. Check what is actually running from inside the head node
sinfo && squeue
Three habits worth building: a CloudWatch billing alarm at 50% of your expected monthly spend, a `MinCount: 0` default you override deliberately rather than accidentally, and a rule that the head node gets stopped (`pcluster stop` stops the fleet, not the head node — terminate the cluster if you won't be back for a week). If keeping track of fleet states across client accounts isn't something you want to do manually, that's part of what we cover at awsgood.com.
## FAQ
**Why did my bill jump when I only ran two jobs?**
Almost always a non-zero `MinCount` or a long `ScaledownIdletime`. Check `pcluster describe-compute-fleet` and `sinfo` for nodes in `idle` (billing) rather than `idle~` (powered down). The second most common cause is a NAT gateway left running after the cluster's networking was orphaned.
**Is FSx for Lustre billed when no jobs run?**
Yes. Lustre is billed on provisioned capacity per GB-month, whether or not anything reads it. A 1.2 TB scratch filesystem is roughly $170/month sitting completely idle — that's why scratch filesystems belong to the cluster lifecycle, deleted with it.
**Do I pay for traffic between compute nodes?**
Within a single AZ on the same cluster, no — that traffic stays on the local network. Cross-AZ traffic is billed per GB in both directions, which is one reason to keep a cluster single-AZ even when capacity is tight.
## Sources
- [AWS ParallelCluster — product overview and pricing statement](https://aws.amazon.com/hpc/parallelcluster/)
- [AWS ParallelCluster User Guide — what the service is and what it provisions](https://docs.aws.amazon.com/parallelcluster/latest/ug/what-is-aws-parallelcluster.html)
- [Amazon EC2 On-Demand pricing](https://aws.amazon.com/ec2/pricing/on-demand/)
- [Amazon FSx for Lustre pricing](https://aws.amazon.com/fsx/lustre/pricing/)
FAQ
Does AWS ParallelCluster charge a license fee?
No. ParallelCluster is a free orchestration layer — CloudFormation templates, a Slurm configuration and the `pcluster` CLI. You pay only for the resources it creates: EC2 head and
Leave a Reply