Cost Allocation

# Cost Allocation

## Overview

HPC resources cost money -- whether it is capital expense for on-prem hardware or pay-per-use cloud instances. Cost allocation answers the question: **who is consuming what, and how do we pay for it?**

Slurm's accounting system tracks every job's resource consumption (CPU hours, GPU hours, memory, wall time) and attributes it to accounts and users. This data feeds chargeback (billing groups for actual costs) and showback (reporting usage without billing) models.

---

## Prerequisites: Slurm Accounting

Cost allocation requires Slurm accounting (slurmdbd) to be enabled and configured with:

1. **Accounts** -- organizational units (departments, labs, grants)
2. **Users** -- mapped to accounts
3. **Associations** -- user-to-account mappings with optional limits

```bash
# Example account hierarchy for a life science organization
sacctmgr add account bioteam Description="BioTeam Organization"
sacctmgr add account structural parent=bioteam Description="Structural Biology"
sacctmgr add account drugdisc parent=bioteam Description="Drug Discovery"
sacctmgr add account genomics parent=bioteam Description="Genomics"

# Add users to accounts
sacctmgr add user jsmith Account=structural
sacctmgr add user kpatel Account=drugdisc DefaultAccount=drugdisc
```

Users can belong to multiple accounts and specify which account to charge at job submission:

```bash
sbatch --account=structural my_cryoem_job.sh
sbatch --account=drugdisc my_glide_job.sh
```

---

## Chargeback vs. Showback

| Model | Description | Billing | Use Case |
|-------|-------------|---------|----------|
| **Showback** | Report usage, no actual billing | Informational | Academic, internal teams |
| **Chargeback** | Bill groups for actual resource consumption | Real money | Core facilities, shared services |

Most life science HPC operations start with showback and evolve toward chargeback as the organization matures.

---

## Usage Reporting with sreport

### CPU Hours by Account

```bash
# Total CPU hours per account, past quarter
sreport cluster AccountUtilizationByUser \
    Start=2026-01-01 End=2026-04-01 -t Hours \
    Format=Account,Login,Used

# Account       Login     Proper Name     Used(CPU Hours)
# ---------- ---------- --------------- ----------
# structural  jsmith     J. Smith           42,500
# structural  mlee       M. Lee             28,300
# drugdisc    kpatel     K. Patel           65,100
# drugdisc    achen      A. Chen            31,200
# genomics    rwilson    R. Wilson          98,400
```

---

## Usage Reporting: GPU and Trends

### GPU Hours by Account

```bash
# GPU hours per account (critical for cryo-EM and ML)
sreport cluster AccountUtilizationByUser \
    Start=2026-01-01 End=2026-04-01 -t Hours \
    -T gres/gpu \
    Format=Account,Login,Used

# Account       Login     Proper Name     Used(GPU Hours)
# ---------- ---------- --------------- ----------
# structural  jsmith     J. Smith           12,800
# structural  mlee       M. Lee              8,200
# drugdisc    kpatel     K. Patel            4,500
# genomics    rwilson    R. Wilson              200
```

---

## Usage Reporting: Top Consumers and Trends

### Top Consumers

```bash
# Top 10 accounts by usage
sreport account TopUsage Start=2026-01-01 End=2026-04-01 TopCount=10

# Top 20 users by usage
sreport user TopUsage Start=2026-01-01 End=2026-04-01 TopCount=20
```

### Monthly Trend Reports

Generate monthly usage for trend analysis:

```bash
for month in 01 02 03; do
    echo "=== 2026-$month ==="
    sreport cluster AccountUtilizationByUser \
        Start=2026-${month}-01 End=2026-$((month+1))-01 -t Hours \
        Format=Account,Used
done
```

---

## Cost Models

### On-Premise: Amortized Cost per Core-Hour

For on-prem clusters, calculate the fully loaded cost per core-hour:

```
Annual cluster cost = Hardware amortization + Power + Cooling + Staff + Facilities
                    = $500,000 + $120,000 + $40,000 + $200,000 + $50,000
                    = $910,000/year

Available core-hours = Cores × Hours/year × Target utilization
                     = 2,048 cores × 8,760 hours × 0.75
                     = 13,468,800 core-hours

Cost per core-hour   = $910,000 / 13,468,800 = $0.068
```

For GPU nodes, calculate separately:

```
GPU node annual cost = $80,000 amortization + $15,000 power
                     = $95,000/year per 4-GPU node

GPU-hours available  = 4 GPUs × 8,760 hours × 0.70 = 24,528 GPU-hours
Cost per GPU-hour    = $95,000 / 24,528 = $3.87
```

---

## Cost Models: Cloud and Mapping

### Cloud: Direct Instance Cost

For ParallelCluster or PCS, the cost is straightforward -- EC2 instance pricing:

| Instance | On-Demand/hr | GPUs | CPU-hour Cost | GPU-hour Cost |
|----------|-------------|------|---------------|---------------|
| c5.4xlarge | $0.68 | - | $0.043 | - |
| r5.12xlarge | $3.024 | - | $0.063 | - |
| g5.12xlarge | $5.672 | 4x A10G | $0.118 | $1.42 |
| p4d.24xlarge | $32.77 | 8x A100 | $0.341 | $4.10 |

Add storage (FSx Lustre, EFS), data transfer, and NAT gateway costs for a complete picture.

### Mapping Slurm Usage to Costs

```bash
# Export usage data for a billing period
sacct -S 2026-03-01 -E 2026-04-01 \
    --format=Account,User,Partition,AllocCPUS,AllocTRES,Elapsed,CPUTimeRAW \
    --parsable2 --noheader > usage_march_2026.csv
```

Then multiply by the cost-per-unit for each resource type:

```
Monthly charge = (CPU-hours × $CPU_rate) + (GPU-hours × $GPU_rate) + (Storage-GB × $storage_rate)
```

---

## Enforcing Budgets with QOS Limits

Slurm can enforce spending caps via QOS and association limits:

### Per-Account Limits

```bash
# Limit the structural biology group to 500 concurrent CPU cores
sacctmgr modify account structural set GrpTRES=cpu=500

# Limit the drug discovery group to 16 concurrent GPUs
sacctmgr modify account drugdisc set GrpTRES=gres/gpu=16

# Limit a single user to 5 concurrent jobs
sacctmgr modify user kpatel set MaxJobs=5
```

### QOS for Tiered Service

```bash
# Standard QOS (default)
sacctmgr add qos standard Priority=50 MaxTRESPerUser=cpu=128,gres/gpu=4

# Premium QOS (for time-sensitive work, higher priority)
sacctmgr add qos premium Priority=100 MaxTRESPerUser=cpu=512,gres/gpu=16

# Burst QOS (for large campaigns, lower priority)
sacctmgr add qos burst Priority=10 MaxSubmitJobsPerUser=1000
```

Assign QOS to accounts:

```bash
sacctmgr modify account structural set QOS=standard,premium
sacctmgr modify account drugdisc set QOS=standard,premium,burst
```

Users select QOS at job submission:

```bash
sbatch --qos=premium --account=drugdisc urgent_fep.sh
```

---

## Cloud Cost Attribution

### ParallelCluster

Tag EC2 instances by Slurm account using a prolog script or custom launch template. Use AWS Cost Explorer with tags to break down spending by group.

### PCS

PCS compute node groups map to partitions. If you design one partition per department (e.g., `cryoem-gpu`, `drugdisc-cpu`), AWS Cost Explorer can attribute costs by instance group.

For shared partitions, extract usage from `sacct` and multiply by the hourly instance cost for the period.

### Spot Instance Savings

Track Spot vs. On-Demand usage:

```bash
# Jobs on spot partition
sacct -S 2026-03-01 --partition=spot-cpu \
    --format=Account,AllocCPUS,Elapsed,CPUTimeRAW --noheader
```

Report the savings: Spot typically costs 60-90% less than On-Demand.

---

## Sample Monthly Report

A monthly cost allocation report might include:

```
=== BioTeam HPC Usage Report: March 2026 ===

Cluster Utilization: 74.2% (target: 70-85%)
Total CPU-hours consumed: 265,500
Total GPU-hours consumed: 25,800

Usage by Account:
  Account         CPU-hours   GPU-hours   Est. Cost
  structural      70,800      21,000      $86,220
  drugdisc        96,300       4,500      $24,120
  genomics        98,400         300       $6,810

Queue Wait Times (average):
  cpu partition:    22 min
  gpu partition:    1.8 hrs
  highmem:          45 min

Recommendations:
  - GPU queue wait times elevated; consider 2 additional A100 nodes
  - Genomics group not using GPUs; CPU allocation adequate
  - Drug discovery burst campaign completed; QOS limits can revert
```

Accounts & Fairshare -- sacctmgr account setup
Capacity Planning -- utilization analysis
Monitoring & Accounting -- reporting tools
AWS ParallelCluster -- cloud cost context

Cost Allocation

Related Modules¶

References¶