Sustained Use Discounts (SUDs): Automatic Savings for Compute Engine

Sustained Use Discounts (SUDs) là công cụ chiết khấu automatic mà Google Cloud áp dụng khi bạn chạy tài nguyên Compute Engine dài hạn. Không cần purchase commitment, không cần paperwork — chỉ cần chạy VMs hoặc persistent disks liên tục.

Vẻ ngoài, SUD nghe có vẻ tuyệt vời: "miễn phí discount". Nhưng trong thực tế, SUD là lựa chọn thứ yếu so với CUD, và hiểu tại sao là hiểu cơ chế pricing của GCP.

---

Internal Model: SUD là Tầng Automatic Discount

Cơ Chế SUD

SUD hoạt động dựa trên cumulative usage trong một tháng. Mỗi tầng usage:

Nguyên tắc: Bạn chạy VMs càng lâu, discount càng cao
  
Tầng 1 (0–25% of month):        Giá on-demand, 0% discount
Tầng 2 (25–50% of month):       3% discount
Tầng 3 (50–75% of month):       10% discount
Tầng 4 (75–100% of month):      30% discount

Ví dụ cụ thể:

Month có 730 giờ (30 ngày × 24 giờ)

Scenario: Chạy n2-standard-4 (on-demand $0.19/giờ)

Case 1: Chạy 100 giờ (13.7% of month)
  → 0% discount, cost = 100 × $0.19 = $19

Case 2: Chạy 365 giờ (50% of month)
  → Tầng 1: 25% × 365 = 91.25 giờ × $0.19 × 100% = $17.34
  → Tầng 2: 25% × 365 = 91.25 giờ × $0.19 × 97% = $16.81
  → Total = $34.15 (5.5% average discount)

Case 3: Chạy 700 giờ (95.9% of month)
  → Tầng 1–3: (~400 giờ) with escalating discounts
  → Tầng 4: (~300 giờ) × $0.19 × 70% = $40
  → Total ≈ $115 (roughly 25% average discount)

Chú ý: Discount tích lũy per resource type — nếu chạy 2 n2-standard-4 VMs, usage cộng lại, không riêng lẻ.

Áp Dụng Scope & Resources

SUD áp dụng cho:

• Compute Engine VMs — on-demand instances chạy dài hạn
• Persistent Disks — GCE storage
• GPUs/TPUs — GPU/TPU resources khi attached to VMs

SUD không áp dụng cho:

• Spot VMs / Preemptible VMs — chúng đã có built-in discount, SUD double-dip không được
• Committed resources — nếu bạn mua CUD, nó replace SUD
• Cloud Run, Cloud Functions — managed services khác, không tính SUD

---

Tại Sao SUD < CUD: Discount Rate Comparison

Discount rates:

SUD (automatic):
  - Tầng 4 (75–100% usage): ~30% discount

CUD (purchased):
  - 1-year resource-based: 55% discount
  - 3-year resource-based: 70% discount
  - 1-year flexible spend: 25–35% discount

Tại sao? CUD là advance commitment — bạn promising Google "tôi sẽ chạy này 1–3 năm", nên Google cho discount cao hơn. SUD là "tôi chạy, bạn tự động discount" — less predictable, lower discount rate.

Kết quả: Nếu bạn commit dài hạn (next 3 years) usage sẽ ≥ 75% every month:

• CUD: 70% discount = best choice
• SUD: 30% discount = leaving 40% savings on the table

---

Stacking & Interaction with Other Discounts

SUD + CUD Stacking

Quy tắc: CUD và SUD không stack. Nếu bạn mua CUD:

• Matched usage nhận CUD discount (55–70%)
• Overage usage nhận SUD discount (up to 30%)
• Not both

Ví dụ:

10 CUDs (n2-standard-4) @ 55% discount
Actual usage: 12 VMs

Matched: 10 VMs × 55% discount = $XXX (CUD rate)
Overage: 2 VMs × 30% discount (SUD for overages) = $YYY (on-demand - 30%)
Total = $XXX + $YYY

SUD + Free Tier

Free tier (e.g., 1 f1-micro VM/month) + SUD:

• Free tier capped at specific amount (e.g., 1 f1-micro = 30 f1-micro-hours/month)
• SUD applied on additional usage beyond free tier

---

Operational Limitations of SUD

Limitation 1: Zero Predictability Over Month Boundary

SUD discount resets every month. Nếu chạy 700 giờ ở tháng 1 (95% utilization, 30% discount), tháng 2 reset:

Month 1: 700 giờ → 30% average discount
Month 2: Lại bắt đầu từ 0%, ngay cả nếu chạy liên tục

Điều này tạo monthly billing variance — cost/VM không uniform. Khó để forecast.

Limitation 2: Regional SUD Doesn't Exist

SUD apply per machine type globally. Không có concept "SUD for us-central1 vs us-east1". Nếu chạy:

• 700 giờ n2-standard-4 ở us-central1
• 0 giờ n2-standard-4 ở us-east1

Discount tính trên tổng 700 giờ across both regions (technically, they're different pricing SKUs, but discount stacks).

Limitation 3: No Commitment Buffer

SUD là reactive — bạn chạy, discount tự động apply. Không có "planning":

• Nếu forecast sai (expected 80% utilization, actual 40%), không có refund
• Nếu forecast sai (expected 40%, actual 80%), không có "commit now, get higher discount retroactively"

---

Cost Impact: SUD vs No Discount vs CUD

Let's quantify:

Scenario: n2-standard-4 VM, on-demand = $0.19/giờ

Annual cost (730 giờ/month × 12 month):

No discount: 730 × 12 × $0.19 = $16,632/year

SUD (80% utilization = avg 25% discount):
  = 730 × 12 × $0.19 × 75% = $12,474/year
  Savings: $4,158/year (25%)

1-year CUD (55% discount):
  CUD cost: 730 × 12 × $0.19 × 45% = $7,485
  Savings: $9,147/year (55%)

3-year CUD (70% discount):
  CUD cost (year 1–3): 730 × 12 × $0.19 × 30% = $4,990/year
  Savings: $11,642/year (70%)

Conclusion: CUD là clear winner để predictable, long-term workload. SUD chỉ valuable nếu CUD không feasible (usage pattern terlalu variable).

---

When to Use SUD

Use SUD nếu:

1. Usage pattern unpredictable — không biết apa akan chạy ở tháng depan
2. Không comfortable committing — prefer flexibility over savings
3. Pilot / experimental workload — testing, might spin down anytime
4. Very short bursts — usage ≤ 50% of month, SUD discount minimal anyway

Don't rely on SUD nếu:

1. Core, steady-state workload — commit to CUD, get 55–70% discount
2. Multi-year cost forecast — SUD unpredictable month-to-month
3. Cost accountability — line-item billing per service, SUD makes cost variance high

---

Monitoring & Transparency

How to Track SUD Benefit

1. Cloud Billing → Reports → Sustained Use Discounts

   • Visual breakdown: tháng nào utilization bao nhiêu, discount tầng nào

2. BigQuery export analysis:

   SELECT
     DATE_TRUNC(usage_start_time, MONTH) as month,
     resource.labels.instance_name,
     COUNT(*) as usage_hours,
     SUM(cost) as total_cost,
     SUM(cost) / COUNT(*) as cost_per_hour
   FROM `project.billing_dataset.gcp_billing_export_v1`
   WHERE service.description = 'Compute Engine'
   GROUP BY month, instance_name
   ORDER BY month DESC, cost_per_hour DESC;

3. Comparing to on-demand baseline:

   Actual cost: $X
   On-demand cost (0% discount): $X / 0.75 (if SUD gave 25% avg discount)
   Implied discount: ($X/0.75 - $X) / ($X/0.75) = 25%

---

Anti-Patterns & Mistakes

Anti-Pattern 1: Relying Solely on SUD for Cost Control

Mistake: "Okay, we'll let SUD automatic discounts handle it."

Problem: 30% discount (max SUD) is not enough for production systems targeting cost efficiency. You need explicit commitment.

Fix: Use SUD as baseline (30% discount floor), layer CUD on top for core workloads.

Anti-Pattern 2: Spinning Down & Up Monthly

Scenario: Workload peaks at month-end, then ramps down for first week of next month.

Cost impact:

Month 1: Utilization = 90%, SUD = 30% discount
Month 2: Utilization = 20%, SUD = 0% discount
Average: 15% discount (vs assuming 30%)

Fix: Use committed-capacity VMs (CUD) for baseline, Spot VMs for burst. Avoids monthly SUD reset penalty.

Anti-Pattern 3: Ignoring Preemptible VM Alternative

Situation: Running non-critical batch jobs, SUD discount 30%.

Alternative: Preemptible VMs = 80% discount (fixed, no monthly variance).

Fix: Use Preemptible / Spot VMs for interruptible workloads, not SUD.

---

SUD in the Broader Cost Strategy

SUD should be viewed as baseline benefit you get "for free" by running long-term. But it's not a cost optimization strategy — cost optimization strategy is:

1. CUD for core, predictable → 55–70% savings
2. Spot VMs for interruptible → 60–91% savings
3. Rightsizing + idle removal → 20–40% savings
4. SUD as floor (30% for anything continuous)

Stacking these gets total savings of 60–80% for well-designed systems.

---

References

• Sustained Use Discounts for Compute Engine
• VM Pricing
• Committed Use Discounts vs Sustained Use Discounts
• Cloud Billing Concepts