muxfys

go get github.com/VertebrateResequencing/muxfys

muxfys is a pure Go library for temporarily in-process mounting multiple different remote file systems or object stores on to the same mount point as a "filey" system. Currently only support for S3-like systems has been implemented.

It has high performance, and is easy to use with nothing else to install, and no root permissions needed (except to initially install/configure fuse: on old linux you may need to install fuse-utils, and for macOS you'll need to install osxfuse; for both you must ensure that 'user_allow_other' is set in /etc/fuse.conf or equivalent).

It has good S3 compatibility, working with AWS Signature Version 4 (Amazon S3, Minio, et al.) and AWS Signature Version 2 (Google Cloud Storage, Openstack Swift, Ceph Object Gateway, Riak CS, et al).

It allows "multiplexing": you can mount multiple different S3 buckets (or sub directories of the same bucket) on the same local directory. This makes commands you want to run against the files in your buckets much simpler, eg. instead of mounting s3://publicbucket, s3://myinputbucket and s3://myoutputbucket to separate mount points and running:

$ myexe -ref /mnt/publicbucket/refs/human/ref.fa -i /mnt/myinbucket/xyz/123/
  input.file > /mnt/myoutputbucket/xyz/123/output.file

You could multiplex the 3 buckets (at the desired paths) on to the directory you will work from and just run:

$ myexe -ref ref.fa -i input.file > output.file

It is a "filey" system ('fys' instead of 'fs') in that it cares about performance and efficiency first, and POSIX second. It is designed around a particular use-case:

Non-interactively read a small handful of files who's paths you already know, probably a few times for small files and only once for large files, then upload a few large files. Eg. we want to mount S3 buckets that contain thousands of unchanging cache files, and a few big input files that we process using those cache files, and finally generate some results.

In particular this means we hold on to directory and file attributes forever and assume they don't change externally. Permissions are ignored and only you get read/write access.

When using muxfys, you 1) mount, 2) do something that needs the files in your S3 bucket(s), 3) unmount. Then repeat 1-3 for other things that need data in your S3 buckets.

Performance

To get a basic sense of performance, a 1GB file in a Ceph Object Gateway S3 bucket was read, twice in a row for tools with caching, using the methods that worked for me (I had to hack minfs to get it to work); units are seconds (average of 3 attempts) needed to read the whole file:

| method         | fresh | cached |
|----------------|-------|--------|
| s3cmd          | 5.9   | n/a    |
| mc             | 7.9   | n/a    |
| minfs          | 40    | n/a    |
| s3fs           | 12.1  | n/a    |
| s3fs caching   | 12.2  | 1.0    |
| muxfys         | 5.7   | n/a    |
| muxfys caching | 5.8   | 0.7    |

Ie. minfs is very slow, and muxfys is about 2x faster than s3fs, with no noticeable performance penalty for fuse mounting vs simply downloading the files you need to local disk. (You also get the benefit of being able to seek and read only small parts of the remote file, without having to download the whole thing.)

The same story holds true when performing the above test 100 times ~simultaneously; while some reads take much longer due to Ceph/network overload, muxfys remains on average twice as fast as s3fs. The only significant change is that s3cmd starts to fail.

For a real-world test, some data processing and analysis was done with samtools, a tool that can end up reading small parts of very large files. www.htslib.org/workflow was partially followed to map fastqs with 441 read pairs (extracted from an old human chr20 mapping). Mapping, sorting and calling was carried out, in addition to creating and viewing a cram. The different caching strategies used were: cup == reference-related files cached, fastq files uncached, working in a normal POSIX directory; cuf == as cup, but working in a fuse mounted writable directory; uuf == as cuf, but with no caching for the reference-related files. The local(mc) method involved downloading all files with mc first, with the cached result being the maximum possible performance: that of running bwa and samtools when all required files are accessed from the local POSIX filesystem. Units are seconds (average of 3 attempts):

| method     | fresh | cached |
|------------|-------|--------|
| local(mc)  | 157   | 40     |
| s3fs.cup   | 175   | 50     |
| muxfys.cup | 80    | 45     |
| muxfys.cuf | 79    | 44     |
| muxfys.uuf | 88    | n/a    |

Ie. muxfys is about 2x faster than just downloading all required files manually, and over 2x faster than using s3fs. There isn't much performance loss when the data is cached vs maximum possible performance. There's no noticeable penalty (indeed it's a little faster) for working directly in a muxfys-mounted directory.

Finally, to compare to a highly optimised tool written in C that has built-in support (via libcurl) for reading from S3, samtools was once again used, this time to read 100bp (the equivalent of a few lines) from an 8GB indexed cram file. The builtin(mc) method involved downloading the single required cram cache file from S3 first using mc, then relying on samtools' built-in S3 support by giving it the s3:// path to the cram file; the cached result involves samtools reading this cache file and the cram's index files from the local POSIX filesystem, but it still reads cram data itself from the remote S3 system. The other methods used samtools normally, giving it paths within the fuse mount(s) created. The different caching strategies used were: cu == reference-related files cached, cram-related files uncached; cc == everything cached; uu == nothing cached. Units are seconds (average of 3 attempts):

| method      | fresh | cached |
|-------------|-------|--------|
| builtin(mc) | 1.3   | 0.5    |
| s3fs.cu     | 4.3   | 1.7    |
| s3fs.cc     | 4.4   | 0.5    |
| s3fs.uu     | 4.4   | 2.2    |
| muxfys.cu   | 0.3   | 0.1    |
| muxfys.cc   | 0.3   | 0.06   |
| muxfys.uu   | 0.3   | 0.1    |

Ie. muxfys is much faster than s3fs (more than 2x faster probably due to much faster and more efficient stating of files), and using it also gives a significant benefit over using a tools' built-in support for S3.

Status & Limitations

The only RemoteAccessor implemented so far is for S3-like object stores.

In cached mode, random reads and writes have been implemented.

In non-cached mode, random reads and serial writes have been implemented. (It is unlikely that random uncached writes will be implemented.)

Non-POSIX behaviours:

• does not store file mode/owner/group
• does not support hardlinks
• symlinks are only supported temporarily in a cached writeable mount: they can be created and used, but do not get uploaded
• atime (and typically ctime) is always the same as mtime
• mtime of files is not stored remotely (remote file mtimes are of their upload time, and muxfys only guarantees that files are uploaded in the order of their mtimes)
• does not upload empty directories, can't rename remote directories
• fsync is ignored, files are only flushed on close

Guidance

CacheData: true will usually give you the best performance. Not setting an explicit CacheDir will also give the best performance, as if you read a small part of a large file, only the part you read will be downloaded and cached in the unique CacheDir.

Only turn on Write mode if you have to write.

Use CacheData: false if you will read more data than can be stored on local disk.

If you know that you will definitely end up reading the same data multiple times (either during a mount, or from different mounts) on the same machine, and have sufficient local disk space, use CacheData: true and set an explicit CacheDir (with a constant absolute path, eg. starting in /tmp). Doing this results in any file read downloading the whole remote file to cache it, which can be wasteful if you only need to read a small part of a large file. (But this is the only way that muxfys can coordinate the cache amongst independent processes.)

Usage

import "github.com/VertebrateResequencing/muxfys"

// fully manual S3 configuration
accessorConfig := &muxfys.S3Config{
    Target:    "https://s3.amazonaws.com/mybucket/subdir",
    Region:    "us-east-1",
    AccessKey: os.Getenv("AWS_ACCESS_KEY_ID"),
    SecretKey: os.Getenv("AWS_SECRET_ACCESS_KEY"),
}
accessor, err := muxfys.NewS3Accessor(accessorConfig)
if err != nil {
    log.Fatal(err)
}
remoteConfig1 := &muxfys.RemoteConfig{
    Accessor: accessor,
    CacheDir: "/tmp/muxfys/cache",
    Write:    true,
}

// or read configuration from standard AWS S3 config files and environment
// variables
accessorConfig, err = muxfys.S3ConfigFromEnvironment("default",
    "myotherbucket/another/subdir")
if err != nil {
    log.Fatalf("could not read config from environment: %s\n", err)
}
accessor, err = muxfys.NewS3Accessor(accessorConfig)
if err != nil {
    log.Fatal(err)
}
remoteConfig2 := &muxfys.RemoteConfig{
    Accessor:  accessor,
    CacheData: true,
}

cfg := &muxfys.Config{
    Mount:     "/tmp/muxfys/mount",
    CacheBase: "/tmp",
    Retries:   3,
    Verbose:   true,
}

fs, err := muxfys.New(cfg)
if err != nil {
    log.Fatalf("bad configuration: %s\n", err)
}

err = fs.Mount(remoteConfig, remoteConfig2)
if err != nil {
    log.Fatalf("could not mount: %s\n", err)
}
fs.UnmountOnDeath()

// read from & write to files in /tmp/muxfys/mount, which contains the
// contents of mybucket/subdir and myotherbucket/another/subdir; writes will
// get uploaded to mybucket/subdir when you Unmount()

err = fs.Unmount()
if err != nil {
    log.Fatalf("could not unmount: %s\n", err)
}

logs := fs.Logs()

Provenance

There are many ways of accessing data in S3 buckets. Common tools include s3cmd for direct up/download of particular files, and s3fs for fuse-mounting a bucket. But these are not written in Go.

Amazon provide aws-sdk-go for interacting with S3, but this does not work with (my) Ceph Object Gateway and possibly other implementations of S3.

minio-go is an alternative Go library that provides good compatibility with a wide variety of S3-like systems.

There are at least 3 Go libraries for creating fuse-mounted file-systems. github.com/jacobsa/fuse was based on bazil.org/fuse, claiming higher performance. Also claiming high performance is github.com/hanwen/go-fuse.

There are at least 2 projects that implement fuse-mounting of S3 buckets:

• github.com/minio/minfs is implemented using minio-go and bazil, but in my hands was very slow. It is designed to be run as root, requiring file-based configuration.
• github.com/kahing/goofys is implemented using aws-sdk-go and jacobsa/fuse, making it incompatible with (my) Ceph Object Gateway.

Both are designed to be run as daemons as opposed to being used in-process.

muxfys is implemented using minio-go for compatibility, and hanwen/go-fuse for speed. (In my testing, hanwen/go-fuse and jacobsa/fuse did not have noticeably difference performance characteristics, but go-fuse was easier to write for.) However, some of its read code is inspired by goofys. Thanks to minimising remote calls to the remote S3 system, and only implementing what S3 is generally capable of, it shares and adds to goofys' non-POSIX behaviours.

Versioning

This project adheres to Semantic Versioning. See CHANGELOG.md for a description of changes.

If you want to rely on a stable API, vendor the library, updating within a desired version. For example, you could use Glide and:

$ glide get github.com/VertebrateResequencing/muxfys#^2.0.0