ZeroFS vs AWS Mountpoint-s3 Benchmarks

Performance comparison conducted on Azure D48lds v6 (48 vCPUs, 96 GiB RAM) with Cloudflare R2 backend. Results published August 2025.

Test Setup

• VM: Azure Standard D48lds v6, West Europe (Zone 1)
• Storage: Cloudflare R2 (S3-compatible)
• Benchmark suite: github.com/Barre/ZeroFS/bench
• Operations per test: 100 (reduced from 10,000 so the Mountpoint-s3 runs would complete in practical time)

Architecture Differences

ZeroFS: Stores file data and metadata in its own layout in the S3 bucket and implements POSIX semantics on top: in-place writes, appends, permissions, timestamps. 8,662 pjdfstest cases run in CI on every change; a few cases per protocol are excluded, and the lists are public in the repo. No additional infrastructure is required.

AWS Mountpoint-s3: Amazon's official S3 FUSE mount (github.com/awslabs/mountpoint-s3). It maps each S3 object 1:1 to a file or directory. In-place modification, appends, and some metadata operations (utime, chmod, chown) are not supported.

Benchmark Results

Synthetic Benchmarks

Test	ZeroFS	AWS Mountpoint-s3	Difference
Sequential Writes
Operations/sec	663.87	0.70	948x
Mean latency	1.42ms	1,435.81ms	1,011x
Success rate	100%	100%	-
Data Modifications
Operations/sec	695.53	N/A	-
Mean latency	1.30ms	N/A	-
Success rate	100%	0%	Not supported
Single File Append
Operations/sec	769.50	N/A	-
Mean latency	1.22ms	N/A	-
Success rate	100%	0%	Not supported
Empty Files
Operations/sec	888.66	0.09	9,874x
Mean latency	0.86ms	605.61ms	704x
Success rate	100%	2%	-
Empty Directories
Operations/sec	985.98	2.08	474x
Mean latency	0.98ms	479.80ms	490x
Success rate	100%	100%	-
Random Reads
Operations/sec	1,000.84	3.20	313x
Mean latency	0.90ms	312.13ms	347x
Success rate	100%	100%	-

Real-World Operations

Operation	ZeroFS	AWS Mountpoint-s3	Notes
Git clone	3.1s	Failed	Configuration file operations not supported
tar -xf (ZFS source)	13.5s	~2h (est.)	Extrapolated from 10% completion at 12m 27s

Key Observations

ZeroFS

• Mean latencies between 0.86ms and 1.42ms across the six synthetic tests
• 100% success rate in every test
• Completed both real-world tasks: git clone in 3.1s, tar extraction in 13.5s

AWS Mountpoint-s3

• Designed for read-heavy workloads with direct S3 object mapping
• Does not support file modification or append operations by design
• No utime, chmod, or chown support

Technical Details

Sequential Writes

Creates files in sequence. Tests metadata performance and write throughput.

ZeroFS: 100 files in 150ms
Mountpoint-s3: 100 files in 143 seconds (948x slower)

Data Modifications

Random writes to existing files. Tests consistency and update capability.

ZeroFS: All operations succeeded
Mountpoint-s3: Not supported by the version tested

Single File Append

Appends to a single file. Tests sequential write patterns.

ZeroFS: All operations succeeded
Mountpoint-s3: Not supported by the version tested

Empty File Creation

Pure metadata operations without data writes.

ZeroFS: 100 files in 112ms
Mountpoint-s3: 2 of 100 operations succeeded

Empty Directory Creation

Tests directory metadata operations.

ZeroFS: 100 directories in 101ms
Mountpoint-s3: 100 directories in 48 seconds (474x slower)

Random Reads

Tests read performance from various file positions.

ZeroFS: 1,000.84 ops/sec
Mountpoint-s3: 3.2 ops/sec (313x slower)

Git Clone

Tests mixed read/write patterns with metadata operations.

ZeroFS: Completed in 3.1 seconds
Mountpoint-s3: Unable to complete due to lack of config file modification support

Archive Extraction

Extracting ZFS 2.3.3 source tarball. Tests file creation with permissions and timestamps.

ZeroFS: 13.5 seconds for complete extraction
Mountpoint-s3: 12 minutes 27 seconds for 10% (432 of 4,280 files)

• Extrapolated ~2 hours for complete extraction
• Permission operations not fully supported

Storage Efficiency

S3 Operations Comparison

Metric	ZeroFS	AWS Mountpoint-s3	Notes
Class A Operations	578	8,770	15.2x more
Class B Operations	61	5,870	96.2x more

Note: Mountpoint-s3 numbers exclude operations for the remaining 90% of tar extraction.

Design Philosophy Differences

AWS Mountpoint-s3 prioritizes:

• Direct S3 object mapping - 1:1 correspondence between S3 objects and files
• Read-optimized access - Designed primarily for reading existing S3 data
• S3 consistency model - Maintains S3's consistency semantics

ZeroFS prioritizes:

• POSIX semantics - In-place modification, append, permissions, and timestamps
• Operation latency - Mean latencies between 0.86ms and 1.42ms in these tests
• General-purpose workloads - Tools that mix reads, writes, and metadata operations, as in the git clone and tar extraction tests

Summary

Mountpoint-s3 maps each S3 object 1:1 to a file. That mapping excludes in-place modification, appends, and some metadata operations (utime, chmod, chown). ZeroFS stores data in its own layout in the bucket and implements POSIX semantics on top, with S3 as the only storage backend. Both systems ran the same workload on the same VM and bucket; the tables above hold the results.