Linux Clusters Institute:�Current HPC File Systems & Trends

J.D. Maloney | Lead HPC Storage Engineer

Storage Enabling Technologies Group (SET)

National Center for Supercomputing Applications (NCSA)

malone12@illinois.edu

University of Oklahoma, May 13^th – 17^th 2024

Baseline

• Look over slides the beginner workshop in Feb 2024
• Link to Slides
• Segments of some following slides were adapted from their work
• Have an grasp on:
• Definition of Parallel file system
• Types of file systems (home, scratch, projects, archive)

2

May 13 – 17, 2024

• There are many file systems out there, too numerous to hit them all
• Going to touch on current HPC file systems that are popular
• Trying to hit the ones you’re most likely to encounter in an HPC environment these days
• Presentation on a given product is by no means an endorsement of any kind for the vendor discussed
• Will wrap up by discussing some HPC file system trends

3

May 13 – 17, 2024

Topic Coverage

Current HPC-Relevant File Systems

4

May 13 – 17, 2024

Storage Scale (GPFS) Overview

• Product of IBM, gone through many name changes
• Tiger Shark FS, MultiMedia FS, GPFS, Spectrum Scale, now its Storage Scale
• Licensed file system, based on capacity (usually)
• One of the two “prominent” file systems in recent years used by some of the world’s largest supercomputers
• Generally considered easier to administer due to product maturity and Enterprise level features
• Can run on top of RAID’d LUNs presented over a SAN, or it can erasure across disks in servers (Erasure Code Edition)

5

May 13 – 17, 2024

Storage Scale (GPFS) Architecture

6

May 13 – 17, 2024

Image Credit: redlineperf.com

Storage Scale (GPFS) Architecture Notes

• Types of servers:
• NSD Servers- Connected to disks, serve I/O to other nodes in the cluster
• Manager Nodes- Cluster Manager, FS manager, Quorum Nodes
• Clients- Mount file system for access, run GPFS daemon also
• Supports multiple storage pools and has HSM functionality
• Supports encryption at rest (Advanced License)
• Features Like:
• AFM (Active File Management)
• Built in Policy Engine (very powerful)
• Native Cluster Export Services NFS, Samba/CIFS
• Support for Object Storage via Cluster Export Services
• Remote Cluster Mounts
• Powerful Sub-block Allocation

7

May 13 – 17, 2024

Lustre Overview

• Open Source file system supported, developed by many companies and large institutions (DDN, Seagate, Intel, CEA, DOE)
• One of the two “prominent” file systems in use today by the world’s largest supercomputers
• Known for its ability to scale sequential I/O performance as the storage system grows
• More complicated to administer, stricter operating environment (OFED stack, kernel, etc.)
• Can grow to greater numbers of clients

8

May 13 – 17, 2024

Lustre Architecture

9

May 13 – 17, 2024

Image Credit: nextplatform.com

Lustre Architecture Notes

• Very saleable historically for OSS/OSTs (data storage/performance), but now also for MDS/MDTs (metadata storage/performance)
• Built in quotas, but no built in Policy Engine…yet (external via Robinhood)
• Supports mixed fabrics within the same environment via LNET routing
• HSM support is built in, can be manipulated with external policy engine to migrate files between tiers
• Can run more ”commodity” storage, but expects reliable disk, HA maintained by failover

10

May 13 – 17, 2024

BeeGFS Overview

• File System formerly known as FgHS (Fraunhofer File System)
• Maintained by the Fraunhofer GmBH
• Most features and abilities are open source, support and some advance features can be purchased through Fraunhofer
• Scaling of both data and metadata performance and capacity as hardware used grows
• More limited features set than other big file systems:
• No encryption support
• No HSM supported (yet)
• Gaining in popularity recently, especially in Europe
• Now has a policy engine (new as of Late 2022)

11

May 13 – 17, 2024

BeeGFS Architcture

12

May 13 – 17, 2024

Image Credit: beegfs.io

BeeGFS Architecture Notes

• Scale metadata and data services independently
• Multiple services can be run from same node
• Host failure resistance by using ”Buddy Groups” to replicated data on a per directory granularity for high uptime
• Runs on “commodity” hardware
• Supportive of multiple fabric types (Ethernet, IB, OPA)
• Storage pools allow grouping of LUNs by different types (SSD, HDD, etc.)

13

May 13 – 17, 2024

Ceph Overview

• Object/Posix file system that is developed by Red Hat gaining a lot of popularity
• Usually leveraged as an “Archive” store, a general dataset repository, and a backing to Openstack
• Open Source, free to use, support available through Red Hat
• Provides Object, Block, and Posix storage
• Very popular in the cloud/virtualization space
• Runs on “commodity” hardware
• Can be deployed over Ethernet or IB
• Ethernet is very popular with this FS

14

May 13 – 17, 2024

Ceph Architecture

15

May 13 – 17, 2024

Image Credit: innotta.com.au

Ceph Architecture Notes

• Uses the CRUSH algorithm to handle file storage and retrieval across the cluster
• Disks are handled individually, each disk has a ”journal” device that allows for quick writes for small files
• Multiple disks can share the same journal device, tuning the ratio and drive types allows one to increase or decrease performance
• Supports both replication or erasure coding for redundancy
• Does NOT expect any devices to be RAID’d together before being presented
• Disks are individually formatted by Ceph (Bluestore) and managed by the file system
• Multi metadata server support (for POSIX component) is maturing

16

May 13 – 17, 2024

VAST Overview

17

May 13 – 17, 2024

• Proprietary file system and hardware solution from VAST Data that talks NFS, SMB, and S3
• Based on all-flash premise, does not support HDDs
• Closed source, need to buy VAST hardware and support
• Supports multiple cluster-side fabrics (IB, Ethernet, others coming)
• Popular as a hands off, low maintenance, highly available file system solution

VAST Architecture

18

May 13 – 17, 2024

VAST Architecture Notes

• All C-Boxes are connected to all D-boxes
• Writes all land on Optane drives (or alternate high endurance drive option as Intel Optane fades)
• This gets all I/O aligned and files ready to be flushed to the back-end QLC capacity drives
• Read performance comes from all drives; but this limits writes to just the performance of the high endurance drives
• Per the above bullet; solution is capable of high read amounts, less so for writes (know your workload)
• Backend network can be either Ethernet or Infiniband (but usually is Ethernet)

19

May 13 – 17, 2024

Others

• There are many other really cool file systems out there to choose from:
• WekaIO
• DAOS
• Quobyte
• HDFS
• List Goes On
• Always good to keep an eye on these as everything has to start somewhere
• Helps you understand which direction things are headed
• Where the heaviest development activity is taking place

20

May 13 – 17, 2024

HPC File System Trends

21

May 13 – 17, 2024

New Ground-Up File Systems

• With the increasing prevalence of flash and its economics of use in more places, new ground-up file systems are being written with flash and/or flash + object in mind
• New technology such as DPUs also plays a roll in accelerating storage I/O
• These new technologies can allow for more feature implementation with much less/no cost penalty
• Things like compression, deduplication, more advanced application/user performance telemetry, etc.
• These file systems are also being written to be much more resilient and fault tolerant, also to be cloud-native

22

May 13 – 17, 2024

New Deployment Methods

• Historical FS deployment has been RPM install or running a fully controlled OS image from a vendor
• Has worked well for the most part (especially full images where all packages/configs are controlled)…but can be brittle and upgrades stink
• Newer file systems are deployed via containers
• Fully encapsulated to ensure packages and libraries are approved versions
• Greatly improves upgrade processes/expansion/contraction, easier to spin up and spin down
• At least partly insulates vendors from a shifting OS landscape and allows sites to run the bare metal OS of their choice

23

May 13 – 17, 2024

Wrap Up

• Many of file systems mentioned in this presentation are open source and/or available to play with for free
• Set some of the appealing ones to you up and test them out in VMs or on older hardware
• Keep up on new releases as features get added or weak points get addressed
• Tons of options out there to choose from
• No one right solution, many factors go into making the decision
• Balance the trade-offs, your environment will have different constraints than someone else’s
• Reach out to others in the community, attend User Group meetings

24

May 13 – 17, 2024

Acknowledgements

• Members of the SET group at NCSA for slide review
• Members of the steering committee for slide review

25

May 13 – 17, 2024

Questions

26

May 13 – 17, 2024