OpenZFS at LinuxCon

LLNL-PRES-643675
This work was performed under the auspices of the U.S. Department
of Energy by Lawrence Livermore National Laboratory under Contract
DE-AC52-07NA27344. Lawrence Livermore National Security, LLC
LinuxCon 2013
Brian Behlendorf, Open ZFS on Linux
September 17, 2013

Lawrence Livermore National Laboratory LLNL-PRES-xxxxxx
2
●
Advanced Simulation
– Massive Scale
– Data Intensive
●
Top 500
– #3 Sequoia
●
20.1 Peak PFLOP/s
●
1,572,864 cores
●
55 PB of storage at 850 GB/s
– #8 Vulcan
●
5.0 Peak PFLOPS/s
●
393,216 cores
●
6.7 PB of storage at 106 GB/s
●
●
High Performance Computing
World class computing resources

Lawrence Livermore National Laboratory LLNL-PRES-643675
3
●
First Linux cluster deployed in 2001
●
Near-commodity hardware
●
Open Source
●
Clustered High Availability Operating
System (CHAOS)
– Modified RHEL Kernel
– New packages – monitoring,
power/console, compilers, etc
– Lustre Parallel Filesystem
– ZFS Filesystem
Linux Clusters
LLNL Loves Linux

4
●
Massively parallel distributed filesystem
●
Lustre servers use a modified ext4
– Stable and fast, but...
– No scalability
– No data integrity
– No online manageability
●
Something better was needed
– Use XFS, BTRFS, etc?
– Write a filesystem from scratch?
– Port ZFS to Linux?
Lustre Filesystem
Existing Linux filesystems do not meet our requirements

5
●
2008 – Prototype to determine viability
●
2009 – Initial ZVOL and Lustre support
●
2010 – Development moved to Github
●
2011 – POSIX layer added
●
2011 – Community of early adopters
●
2012 – Production usage of ZFS
●
2013 – Stable GA release
ZFS on Linux History
Community involvement was exceptionally helpful

6
ARC
ZIO
VDEV Configuration
DMU
ZAP
ZIL
DSL
Traversal
OST
OFD
ZFS OSD
MDT
MDD
ZVOL /dev/zfsZPL
libzfs
ZFS CLI
Interface
Layer
Transactional
Object
Layer
Pooled
Storage
Layer
User
Kernel
Architecture
SPL

Lawrence Livermore National Laboratory LLNL-PRES-xxxxxx
7
●
Core ZFS code is self contained
– May be built in user space or kernel space
– Includes functionality for snapshots, clones, I/O pipeline, etc
●
Solaris Porting Layer
– Adds stable Solaris/Illumos interfaces
●
Taskqs, lists, condition variables, rwlocks, memory allocators, etc...
– Layered on top of Linux equivalents if available
– Solaris specific interfaces were implemented from scratch
●
Vdev disk
– Interfaces with the Linux kernel block layer
– Had to be rewritten to use native Linux interfaces
Linux Specific Changes
The core ZFS code required little change

8
●
Interface Layer
– /dev/zfs
●
Device node interface for user space zfs and zpool utilities
●
Minor changes needed for a Linux character device
– ZVOL: ZFS Volumes
●
Reimplemented as Linux block driver which is backed by the DMU
– ZPL: ZFS Posix Layer
●
Most complicated part, there are significant VFS differences
●
In general new functions were added for the Linux VFS handlers
●
If possible the Linux handlers use the equivalent Illumos handler
– Lustre
●
Support for Lustre was added by Sun/Oracle/Whamcloud/Intel
●
Lustre directly layers on the DMU and does not use the Posix Layer
Linux Specific Changes
The majority of changes to ZFS were done in the interface layer

9
●
8k stacks
– Illumos allows larger stacks
– Needed to get stack usage down to support stock distribution kernels
– Code reworked as needed to save stack
●
Gcc
– ZFS was written for C99, Linux kernel is C89
– Fix numerous compiler warnings
●
GPL-only symbols
– ZFS is under an open source license but may not use all exported symbols
– This includes basic functionality such as work queues
– ZVOLs can't add entries in /sys/block/
– .zfs/snapshot's can't use the automounter
●
User space
– Solaris threads used instead of pthreads
– Block device naming differences
– Udev integration
Porting Issues

10
●
Memory management
– ZFS relies heavily on virtual memory for its data buffers
●
By design the Linux kernel discourages the use of virtual memory
●
To resolve this the SPL provides a virtual memory based slab
– This allows use of the existing ZFS IO pipeline without modification
– Fragmentation results in underutilized memory
– Stability concerns under low memory conditions
●
We plan to modify ZFS to use scatter-gather lists of pages under Linux
●
Allows larger block sizes
●
Allows support for 32-bit systems
– The ARC is not integrated with the Linux page cache
●
Memory used by the ARC is not reported as cached pages
●
Complicates reclaiming memory from the ARC when needed
●
Requires an extra copy of the data for mmap'ed I/O
Porting Issues

11
●
Features
– O_DIRECT
– Asynchronous IO
– POSIX ACLs
– Reflink
– Filefrag
– Fallocate
– TRIM
– FMA Infrastructure (event daemon)
– Multiple Modified Protection (MMP)
– Large blocks
Future Work
Possibilities for future work

12
●
All source code and the issue tracker are kept at Github
– http://github.com/zfsonlinux/zfs
– 70 contributors, 171 forks, 816 watchers
●
Not in mainline kernel
– Similar to resier4, unionfs, lustre, ceph, …
– Autotools used for compatibility
– Advantages
●
One code base used for 2.6.26 - 3.11 kernels
●
Users can update ZFS independently from the kernel
●
Simplifies keeping in sync with Illumos and FreeBSD
●
User space utilities and kernel modules can share code
– Disadvantages
●
Support for the latest kernel lags slightly
●
Higher maintenance and testing burden for developers
Source Code

13
●
Stable and used in production
●
Performance is comparable to existing Linux filesystems
●
All major ZFS features are available
– Simplified administration - Stripes, Mirrors, and RAIDZ[1,2,3]
– Online management - ZFS Intent Log (ZIL)
– Snapshots / Clones - L2ARC Tiered Caching
– Special .zfs directory - Transparent compression
– Send/receive of snapshots - Transparent deduplication
– Virtual block devices (ZVOL)
●
Currently used by Supercomputers, Desktops, NAS appliances
●
Enthusiastic user community
– zfs-discuss@zfsonlinux.org
Where is ZFS on Linux Today
ZFS is available on Linux today!

14
●
ZFS is Open Source under the CDDL
●
ZFS is NOT a derived work of Linux
– “It would be rather preposterous to call the Andrew FileSystem a 'derived
work' of Linux, for example, so I think it's perfectly OK to have a AFS module,
for example.”
●
Linus Torvalds
– “Our view is that just using structure definitions, typedefs, enumeration
constants, macros with simple bodies, etc., is NOT enough to make a
derivative work. It would take a substantial amount of code (coming from
inline functions or macros with substantial bodies) to do that.”
●
Richard Stallman
Licensing
ZFS can be used in Linux

15

OpenZFS at LinuxCon

More Related Content

What's hot (20)

Similar to OpenZFS at LinuxCon (20)

Recently uploaded (20)

OpenZFS at LinuxCon