CPU
QEMU Virtual CPU version 2.5+
Memory
7.7Gi (5.9Gi avail)
Storage
96G NVMe (43G free)
Single-Thread
426 events/s
sysbench prime to 20,000 · 1 thread · 10s
Multi-Thread (4 cores)
1,637 events/s
96% of perfect 4× scaling · 10s
CPU Steal Time
1.26%
Share of time the hypervisor gave this VM's CPU to someone else, measured over 60s of full load
~0% = cores are not shared with other tenants
Consistency (10 runs)
0.74% variation
Mean 1,642 events/s · min 1,614 · max 1,664
First → last run: -0.62%
Variation 0.74%, steal 1.26% during this run.
All-core throughput per 6-second run (events/s), bars start at zero
Why this matters: We ran the same all-core workload 10 times back to back and measured throughput each time. On an oversubscribed host, other tenants' load shows up as CPU steal time and uneven runs. Burstable instances (AWS t3/t4g in standard mode, Azure B-series) are also throttled to a fraction of a core once their CPU credits run out. Flat bars and ~0% steal mean the cores are yours.
Sequential Write
356 MB/s
fio 1 MiB blocks, QD32, direct I/O, 15s
Sequential Read
1,420 MB/s
fio 1 MiB blocks, QD32, direct I/O, 15s
Random 4K Read
93,672 IOPS
fio randread, 4 jobs × QD64, 30s
Avg latency: 2724.8 µs
AWS EBS gp3: 3,000 IOPS free / 16,000 max (+$65/mo)
Random 4K Write
33,454 IOPS
fio randwrite, 4 jobs × QD64, 30s
Avg latency: 7623.2 µs
AWS EBS gp3: 3,000 IOPS free / 16,000 max (+$65/mo)
Mixed Random R/W (70/30)
54,792 IOPS
Read 38,313 + write 16,479
Closest to a typical database I/O pattern
| Metric | 639Cloud (NVMe included) | AWS EBS gp3 (baseline) | AWS EBS gp3 (max provisioned) |
|---|
| Random 4K Read IOPS | 93,672 | 3,000 | 16,000 (+$65/mo) |
| Random 4K Write IOPS | 33,454 | 3,000 | 16,000 (+$65/mo) |
| Sequential Read | 1,420 MB/s | 125 MB/s | 1,000 MB/s (+$40/mo) |
| Storage Cost | Included in VM price | $0.08/GB/mo + IOPS | $0.08/GB + $0.005/IOPS + throughput |
Key takeaway: 639Cloud NVMe is included at no extra cost. AWS EBS charges separately for volume size, provisioned IOPS and throughput. Getting to 16,000 IOPS on gp3 adds about $65/mo for storage performance alone, on top of compute.
Transactions Per Second
3,093 TPS
pgbench TPC-B-like, 10 clients, 4 threads, 60s
Scale factor 20 (~300 MB database)
Real SELECT/UPDATE/INSERT transactions, not synthetic I/O
Transaction Latency
3.233 ms avg
Per transaction, 10 concurrent connections
Total Transactions
185,334
Completed in 60 seconds
Zero failed transactions
What pgbench measures: PostgreSQL's own standard benchmark: short read/write transactions against indexed tables, with every commit written to disk. It combines CPU, memory and NVMe latency in the way a real OLTP database uses them.
P95 Response Time
70.0 ms
Target under 200 ms: passed
P50 3.1 ms · P99 81.4 ms · max 291.9 ms
Error Rate
0.00%
Every one of 12,036 requests succeeded
Up to 100 concurrent users
Traffic Served
12,036 requests
57.0 req/s average over 3m31s
Paced load (0.5–1.5s think time per user), so this reflects traffic sent, not the VM's limit
| Operation | Avg Latency | What It Tests |
|---|
| Product Lookups | 3.6 ms | Redis cache + PostgreSQL reads |
| Search Queries | 4.2 ms | PostgreSQL pattern search (ILIKE) |
| Order Writes | 5.7 ms | PostgreSQL INSERT, committed to NVMe |
| CPU Endpoint | 13.5 ms | Prime calculation while under load |
What this test does: Up to 100 virtual users hit a real Node.js + PostgreSQL + Redis stack for 3m31s: catalog reads (35%), search (20%), single lookups (20%), order writes (15%), CPU work (5%) and aggregation queries (5%).
P95 of 70.0 ms means 95% of requests at peak load came back within 70.0 milliseconds.
| Benchmark | Result | Why It Matters |
|---|
| CPU Steal Time | 1.26% | Cores not shared with other tenants |
| CPU Consistency | 0.74% variation over 10 runs | No throttling, no credit system |
| Random 4K Read IOPS | 93,672 | vs AWS EBS gp3 baseline of 3,000 |
| Random 4K Write IOPS | 33,454 | NVMe included, no extra storage charges |
| PostgreSQL TPS | 3,093 | Real database performance on NVMe |
| App P95 Latency | 70.0 ms | Full stack at up to 100 concurrent users |
| App Error Rate | 0.00% | 12036 requests |
| VM Cost | $45/month | vs AWS m5.xlarge ($140) + gp3 storage and IOPS ($8–73) = $148–213/month |
This entire benchmark ran on a single 639Cloud Medium VM at $45/month.
Dedicated CPU + NVMe storage + bandwidth included: predictable performance at roughly half the cost of a comparable AWS setup.
Powered by 100% renewable solar energy.
