cunoFS is a Faster and Cheaper Alternative to FSx for Lustre

FSx for Lustre is a distributed parallel filesystem that’s used as part of HPC (high-performance computing) workloads across a variety of complex fields (including life sciences, machine learning, meteorology, finance, and media). It is AWS’s managed version of the Lustre filesystem, and while it’s a popular high-performance storage option, it can quickly become unaffordable as HPC projects scale.

This article directly compares the cost and performance for FSx for Lustre and its alternatives, as well as discussing their scalability and compatibility with your existing HPC workloads.

cunoFS is a high-performance alternative to Amazon’s FSx for Lustre filesystem

This article compares FSx for Lustre, AWS Mountpoint, and cunoFS. cunoFS allows HPC users to use Amazon’s S3 object storage as a high-performance POSIX-compatible filesystem that outperforms FSx for Lustre on both cost and performance, as demonstrated in the benchmarks on this page. 

cunoFS allows you to run your existing applications and scripts (Python, C++, bash, etc.) as normal, which massively improves your development efficiency. 

Features and compatibility: comparing FSx for Lustre, AWS Mountpoint, and cunoFS

FSx for Lustre vs. Mountpoint for Amazon S3

Mountpoint for Amazon S3 allows you to mount an Amazon S3 bucket as a local filesystem, but it has very limited POSIX compatibility and does not support many common filesystem operations. While Mountpoint is free to use and you only pay for your S3 storage costs, the downside is that it doesn’t work for anything beyond listing files, reading them, or creating them.

Modifying and renaming files, deleting directories, and using symlinks and file locking are all beyond Mountpoint’s capabilities, causing the majority of HPC workloads that try to use it to fail.

FSx for Lustre with S3 integration

FSx for Lustre has an integration with S3 that allows you to sync your FSx for Lustre files with S3 objects. You can set this up by adding a data repository association (a link to your S3 bucket) when creating your FSx for Lustre filesystem. 

This allows you to temporarily import all your data from S3, run your workload on the POSIX-compatible FSx for Lustre filesystem, and then shut down the FSx for Lustre filesystem and export your entire output dataset back to S3. 

Many choose FSX for Lustre with S3 integration when they need better performance than S3 offers natively but do not want to bear the full costs of permanently storing data in FSX for Lustre. Using this will give the same performance as FSx for Lustre does; however, it’s still nowhere near as cheap as running all your workloads on S3.

FSx for Lustre vs. cunoFS

Whether you use the standard FSx for Lustre, or FSx for Lustre with S3 integration, it’s still much more expensive than using cunoFS, which runs all your workloads directly on S3. 

S3 is cheaper than FSx for Lustre, with faster performance, and the ability to scale to larger sizes of workload. However, its weakness has always been its lack of POSIX compatibility. Having to rewrite all your code so that every filesystem call works with S3’s API is tedious and time-consuming.

cunoFS allows you to get around S3’s limitations. Your workloads can continue to make POSIX calls while also using S3, as cunoFS automatically intercepts all filesystem calls and replaces them with the correct S3 API calls.

Venn diagram comparing S3, FSx for Lustre (with and without S3 integration) and cunoFS.
S3 is cheap and high-performance but not POSIX compatible. FSx for Lustre (with and without S3 integration) is POSIX compatible but expensive. cunoFS provides the best of both worlds: high performance and POSIX compatibility for a good price.

Performance benchmarks: cunoFS vs. FSx for Lustre

The table below compares the cost vs. performance (in seconds) of FSx for Lustre and cunoFS. We provide this data based on one copy operation per day for a month and on ten copy operations per day for a month, so you can see how each solution scales.

Metrics based on one copy operation per day for a month

Read Write
Copy 5 x 32 GiB into RAM disk Copy Linux kernel source into RAM disk    Copy 5 x 32 GiB into storage system   Copy Linux kernel source into storage system   
FSx for Lustre (all options)

Duration

(seconds)

278-525s 224-292s 287-1046s 145-186s

Monthly cost

($)

$150-644
FSx for Lustre

SSD Scratch

Duration

(seconds)

525s 245s 1046s 182s

Monthly cost

($)

$150
FSx for Lustre

SSD 125 MBps

Duration

(seconds)

327s 224s 330s 149s

Monthly cost

($)

$178
FSx for Lustre

SSD 250 MBps Duration is interpolated.

Duration

(seconds)

320s 228s 324s 153s

Monthly cost

($)

$225
FSx for Lustre

SSD 500 MBps Duration is interpolated.

Duration

(seconds)

306s 234s 312s 162s

Monthly cost

($)

$365
FSx for Lustre

SSD 1000 MBps

Duration

(seconds)

278s 248s 287s 178s

Monthly cost

($)

$644
FSx for Lustre

HDD 12 MBps

Duration

(seconds)

375s 245s 439s 186s

Monthly cost

($)

$154
FSx for Lustre

HDD 12 MBps with SSD cache

Duration

(seconds)

309s 251s 370s 150s

Monthly cost

($)

$252
FSx for Lustre

HDD 40 MBps

Duration

(seconds)

382s 234s 420s 145s

Monthly cost

($)

$153
FSx for Lustre

HDD 40 MBps with SSD cache

Duration

(seconds)

293s 216s 377s 130s

Monthly cost

($)

$182
cunoFS with S3 *

*Does not include license fee for commercial use

Duration

(seconds)

29s 18s 37s 79s

Monthly cost

($)

$25 $26 $26 $41

The data in this table clearly demonstrates that cunoFS is cheaper and faster than FSx for Lustre when reading and writing data.

Metrics based on ten copy operations per day for a month

This data is based on estimates that were extrapolated from the costs of the single copy per day data, taking into account the cost per GB of each system, and burst credits where relevant. It also assumes that the storage system is idle for the full 24 hours between each run.

Read Write
Copy 5 x 32 GiB into RAM disk Copy Linux kernel source into RAM disk    Copy 5 x 32 GiB into storage system   Copy Linux kernel source into storage system   
FSx for Lustre (all options)

Duration

(seconds)

2776-5252s 2242-2925s 2873-10461s 1445-3769s

Monthly cost

($)

$150-644
FSx for Lustre SSD Scratch

Duration

(seconds)

5253s 2448s 10461s 1822s

Monthly cost

($)

$150
FSx for Lustre SSD 125 MBps

Duration (seconds)

3266s 2242s 3298s 1489s

Monthly cost

($)

$178
FSx for Lustre SSD 250 MBps Duration is interpolated.

Duration

(seconds)

3196s 2275s 3237s 1531s

Monthly cost

($)

$225
FSx for Lustre SSD 500 MBps Duration is interpolated.

Duration

(seconds)

3056s 2342s 31165s 1616s

Monthly cost

($)

$365
FSx for Lustre SSD 1000 MBps

Duration

(seconds)

2776s 2477s 2873s 1784s

Monthly cost

($)

$644
FSx for Lustre HDD 12 MBps

Duration

(seconds)

3750s 2450s 4388s 1863s

Monthly cost

($)

$154
FSx for Lustre HDD 12 MBps with SSD cache

Duration

(seconds)

3088s 2513s 3696s 1503s

Monthly cost

($)

$252
FSx for Lustre HDD 40 MBps

Duration

(seconds)

3820s 2392s 4198s 1445s

Monthly cost

($)

$153
FSx for Lustre HDD 40 MBps with SSD cache

Duration

(seconds)

2925s 2163s 3769s 1298s

Monthly cost

($)

$182
cunoFS with S3*

*Does not include license fee for commercial use

Duration

(seconds)

288s 175s 373s 792s

Monthly cost

($)

$27 $34 $35 $191
Read Write
Copy 5 x 32 GiB into RAM disk Copy Linux kernel source into RAM disk    Copy 5 x 32 GiB into storage system   Copy Linux kernel source into storage system   
FSx for Lustre SSD Scratch

Duration

(seconds)

5253s 2448s 10461s 1822s

Monthly cost

($)

$150
FSx for Lustre SSD 125 MBps

Duration (seconds)

3266s 2242s 3298s 1489s

Monthly cost

($)

$178
FSx for Lustre SSD 250 MBps Duration is interpolated.

Duration

(seconds)

3196s 2275s 3237s 1531s

Monthly cost

($)

$225
FSx for Lustre SSD 500 MBps Duration is interpolated.

Duration

(seconds)

3056s 2342s 31165s 1616s

Monthly cost

($)

$365
FSx for Lustre SSD 1000 MBps

Duration

(seconds)

2776s 2477s 2873s 1784s

Monthly cost

($)

$644
FSx for Lustre HDD 12 MBps

Duration

(seconds)

3750s 2450s 4388s 1863s

Monthly cost

($)

$154
FSx for Lustre HDD 12 MBps with SSD cache

Duration

(seconds)

3088s 2513s 3696s 1503s

Monthly cost

($)

$252
FSx for Lustre HDD 40 MBps

Duration

(seconds)

3820s 2392s 4198s 1445s

Monthly cost

($)

$153
FSx for Lustre HDD 40 MBps with SSD cache

Duration

(seconds)

2925s 2163s 3769s 1298s

Monthly cost

($)

$182
cunoFS with S3*

*Does not include license fee for commercial use

Duration

(seconds)

288s 175s 373s 792s

Monthly cost

($)

$27 $34 $35 $191
Again, cunoFS beats FSx for Lustre on price and performance. There is one metric where FSx for Lustre appears to perform better: the cost of writing many small files can be slightly cheaper. But to make this saving, you’d have to choose one of the cheaper FSx for Lustre instances, and the performance is so much worse that this isn’t worth it just to make a small saving.

You can also see our benchmarks that show how cunoFS outpaces previous solutions like s3fs, goofys, and s3ql.

How to switch from FSx for Lustre to cunoFS

Setting up cunoFS to replace FSx for Lustre takes only a few steps. It’s easy to do, and worthwhile as it’s the best available alternative to FSx for Lustre.

You’ll need to create an S3 bucket and IAM credentials to access it, and then download and install cunoFS. Once it’s connected to your bucket, you can start using it immediately in either Direct Intercept mode (for fast CLI filesystem access) or in our high performance Fusion or high-compatibility Flexmount modes.

Next, copy your data from FSx for Lustre to S3 using cunoFS on the command line, and run your unmodified workloads on it. We recommend using Fusion mode for the highest performance, which combines EFS with S3 via cunoFS for the fastest possible performance, without the increased costs of a pure-EFS deployment. 

The advantages of cunoFS go beyond cost and performance: we don’t scramble files, and we don’t lock you into using our product or a single vendor’s object storage. cunoFS is free for personal use and evaluation, and business users can contact us to discuss our commercially supported plans. We’re also keen to hear about how we can extend cunoFS to meet your needs, so if you have any unique requirements, please don’t hesitate to book a meeting.

Related Posts

cunoFS is a Cheaper and Faster Alternative to FSx for Lustre
Table of Contents