Benchmarking OCI Always Free ARM64 Instance Performance

Introduction

Oracle Cloud Always Free offers ARM64 architecture Ampere A1 Compute instances at no cost. Since CPU cores and memory are available permanently free up to a certain limit, many people use them for personal projects or as CI/CD execution platforms.

In this article, I measure the CPU, cryptographic processing, memory, disk I/O, and network speed of this ARM64 instance using stress-ng, fio, openssl speed, and speedtest, running each test 3 times and summarizing the averages.

Test Environment and Tools

The OS is Oracle Linux 9.6 with an aarch64 (ARM64) kernel.

$ uname -a
Linux oci-free 6.12.0-1.23.3.2.el9uek.aarch64 #1 SMP ... aarch64 GNU/Linux

$ lscpu | grep -E 'Model name|CPU\(s\)'
CPU(s):   4
Model name: Neoverse-N1

The tools used and their purposes are as follows.

stress-ng and fio can be obtained via the EPEL repository.

sudo dnf install -y epel-release
sudo dnf install -y stress-ng fio

Each item was measured 3 times, and the average values are used.

CPU Performance

Using the matrixprod (matrix multiplication) stressor in stress-ng, I applied load for 30 seconds and measured bogo ops/s.

# Single core
stress-ng --cpu 1 --cpu-method matrixprod --metrics-brief --timeout 30s

# All cores (4)
stress-ng --cpu 4 --cpu-method matrixprod --metrics-brief --timeout 30s

stress-ng: metrc: cpu  1609  30.09  30.06  0.00  53.48  53.53
stress-ng: metrc: cpu  6401  30.07  119.64  0.04  212.90  53.49

CPU benchmark results using stress-ng matrixprod (3-run average)

The scaling efficiency from single core to all cores is nearly 4x, indicating almost no inter-core interference. Since Neoverse-N1 is designed to scale core count without sacrificing single-thread performance, this result is expected.

Note that bogo ops/s is a pseudo-metric specific to stress-ng and cannot be compared across different stressors. Here, the purpose is to verify relative scaling efficiency using the same stressor (matrixprod).

Cryptographic Performance

I measured AES-256-CBC and SHA-256 throughput using openssl speed. Neoverse-N1 includes hardware AES instructions (ARMv8 Cryptography Extensions), which significantly accelerate performance compared to software implementations.

openssl speed aes-256-cbc sha256

sha256           72255.22k   261173.17k   740627.88k  1369694.87k  1817938.60k  1869770.57k
aes-256-cbc     834545.80k  1378167.23k  1615928.85k  1677753.69k  1700137.64k  1702980.27k

Each column in the output corresponds to a block size (16B / 64B / 256B / 1KB / 8KB / 16KB).

AES-256-CBC at approximately 1.71 GB/s for 16KB blocks represents a throughput level where encryption processing for HTTPS or VPN can be performed at virtually no cost.

Memory Performance

Using the --vm stressor in stress-ng, I repeatedly read and wrote to a 4GB working set and measured throughput.

stress-ng --vm 1 --vm-bytes 4G --metrics-brief --timeout 30s

stress-ng: metrc: vm  1324315  30.23  26.54  3.65  43810.44  43868.81

One of the 3 runs produced a lower value (38,748 bogo ops/s), likely due to contention with other processes, but the remaining 2 runs were stable in the 44,000–46,000 range.

Disk I/O Performance

I measured sequential and random I/O on the /home volume (100GB) using fio.

# Sequential write
fio --name=seq_write --directory=/home/opc/fio_test \
    --rw=write --bs=1M --size=2G --numjobs=1 \
    --time_based --runtime=30 --iodepth=16 --ioengine=libaio --direct=1 --group_reporting

# Sequential read
fio --name=seq_read --directory=/home/opc/fio_test \
    --rw=read --bs=1M --size=2G --numjobs=1 \
    --time_based --runtime=30 --iodepth=16 --ioengine=libaio --direct=1 --group_reporting

# Random 4K read/write
fio --name=rand_rw --directory=/home/opc/fio_test \
    --rw=randrw --bs=4k --size=1G --numjobs=4 \
    --time_based --runtime=30 --iodepth=32 --ioengine=libaio --direct=1 --group_reporting

# Sequential write
WRITE: bw=51.0MiB/s (53.5MB/s), io=2049MiB, run=40179msec

# Sequential read
READ: bw=46.9MiB/s (49.2MB/s), io=1411MiB, run=30059msec

# Random 4K
read: IOPS=4071, BW=15.9MiB/s
write: IOPS=4094, BW=16.0MiB/s

Disk I/O benchmark results using fio (3-run average)

Sequential throughput was approximately 50 MiB/s, and random 4K IOPS was approximately 4,000.

Why These Numbers

OCI Block Volume performance on the “Balanced” tier (default) scales proportionally with volume size.

Comparing our measured values against the limits:

This is not a “slow environment” but rather a state where the Balanced storage performance limit is being fully utilized. For reference, here is a comparison with block storage from other cloud providers:

Throughput is about 1/3 of AWS gp3, but this may not be an issue depending on the use case.

Network Speed

I used Ookla’s official speedtest CLI, connecting to a fixed server in Tokyo (IPA CyberLab 400G, id: 48463). Using a fixed target server eliminates result variability caused by server selection.

speedtest --server-id=48463 --accept-license --accept-gdpr

Server: IPA CyberLab 400G - Tokyo (id: 48463)
   ISP: Oracle Cloud
Idle Latency:  1.76 ms
    Download:  3829.18 Mbps
      Upload:  3977.41 Mbps
 Packet Loss:  0.0%

Network speed measurement results using speedtest (IPA CyberLab 400G Tokyo) (3-run average)

Download was stable across all 3 runs at 3,790–3,860 Mbps, confirming that OCI’s network bandwidth delivers approximately 3.8 Gbps. Upload showed some variation between measurements but averaged over 3.5 Gbps.

Results Summary

Conclusion

Here is a summary of the OCI Always Free ARM64 instance benchmark results:

• CPU: Per-core throughput is modest, but the 4-core scaling efficiency is nearly 100%, making it well-suited for parallel processing
• Cryptographic processing: Hardware AES instructions deliver 1.71 GB/s for AES-256-CBC, making it suitable for TLS termination and encrypted storage use cases
• Disk: Sequential ~50 MiB/s and random 4K ~4,000 IOPS represent the Balanced tier limits, which is a standard level for cloud block storage
• Network: Fixed-server testing recorded 3.8 Gbps download speed — an impressive bandwidth for a free tier, more than sufficient for large data transfers and high-traffic web services

For a completely free environment, the performance is more than adequate, and it can be utilized for a wide range of purposes including lightweight web servers, scheduled batch processing, and data analysis platforms.

References

クラウド・サービスに無料でアクセス

Oracle Cloud Free Tierは無期限で使用できるAlways Freeクラウド・サービスを企業に提供します。

www.oracle.com

Oracle Cloud Infrastructure Arm Compute

Oracle Cloud Infrastructure Arm Computeでコンピューティング集約型のワークロードを移行、構築、実行します。

www.oracle.com

1. fio - Flexible I/O tester rev. 3.41 — fio 3.41-49-gde3d documentation

fio.readthedocs.io