TASTE-OF-IT

GlusterFS Release 3.10.0 mit neuen Features mehr Performance und Bugfixes

JARVIS
Gluster Logo

Gluster Logo

Das Open-Source verteilte Dateisystem, GlusterFS, wurde vor kurzem mit dem Meilenstein 3.10.0 veröffentlicht. Neben über 315 Bugfixes wurden auch die Performance und Usability verbessert.

GlusterFS 3.10.0 Features und Änderungen

Nachfolgend Stichpunktartig ein Überblick über die Änderungen im neuen Release.

GlusterFS 3.10.0 Release Notes

Major changes and features

Brick multiplexing

Notes for users: Multiplexing reduces both port and memory usage. It does not improve performance vs. non-multiplexing except when memory is the limiting factor, though there are other related changes that improve performance overall (e.g. compared to 3.9).

Multiplexing is off by default. It can be enabled with

# gluster volume set all cluster.brick-multiplex on

Limitations: There are currently no tuning options for multiplexing – it’s all or nothing. This will change in the near future.

Known Issues: The only feature or combination of features known not to work with multiplexing is USS and SSL. Anyone using that combination should leave multiplexing off.

Support to display op-version information from clients

Notes for users: To get information on what op-version are supported by the clients, users can invoke the gluster volume status command for clients. Along with information on hostname, port, bytes read, bytes written and number of clients connected per brick, we now also get the op-version on which the respective clients operate. Following is the example usage:

# gluster volume status <VOLNAME|all> clients

Limitations:

Known Issues:

Support to get maximum op-version in a heterogeneous cluster

Notes for users: A heterogeneous cluster operates on a common op-version that can be supported across all the nodes in the trusted storage pool. Upon upgrade of the nodes in the cluster, the cluster might support a higher op-version. Users can retrieve the maximum op-version to which the cluster could be bumped up to by invoking the gluster volume get command on the newly introduced global option, cluster.max-op-version. The usage is as follows:

# gluster volume get all cluster.max-op-version

Limitations:

Known Issues:

Support for rebalance time to completion estimation

Notes for users: Users can now see approximately how much time the rebalance operation will take to complete across all nodes.

The estimated time left for rebalance to complete is displayed as part of the rebalance status. Use the command:

# gluster volume rebalance <VOLNAME> status

Limitations: The rebalance process calculates the time left based on the rate at while files are processed on the node and the total number of files on the brick which is determined using statfs. The limitations of this are:

Known Issues: As glusterfs does not stored the number of files on the brick, we use statfs to guess the number. The .glusterfs directory contents can significantly skew this number and affect the calculated estimates.

Separation of tier as its own service

Notes for users: This change is to move the management of the tier daemon into the gluster service framework, thereby improving it stability and manageability by the service framework.

This has no change to any of the tier commands or user facing interfaces and operations.

Limitations:

Known Issues:

Statedump support for gfapi based applications

Notes for users: gfapi based applications now can dump state information for better trouble shooting of issues. A statedump can be triggered in two ways:

  1. by executing the following on one of the Gluster servers,
   # gluster volume statedump <VOLNAME> client <HOST>:<PID>

All statedumps (*.dump.* files) will be located at the usual location, on most distributions this would be /var/run/gluster/.

Limitations: It is not possible to trigger statedumps from the Gluster CLI when the gfapi application has lost its management connection to the GlusterD servers.

GlusterFS 3.10 is the first release that contains support for the new glfs_sysrq() function. Applications that include features for debugging will need to be adapted to call this function. At the time of the release of 3.10, no applications are known to call glfs_sysrq().

Known Issues:

Disabled creation of trash directory by default

Notes for users: From now onwards trash directory, namely .trashcan, will not be be created by default upon creation of new volumes unless and until the feature is turned ON and the restrictions on the same will be applicable as long as features.trash is set for a particular volume.

Limitations: After upgrade for pre-existing volumes, trash directory will be still present at root of the volume. Those who are not interested in this feature may have to manually delete the directory from the mount point.

Known Issues:

Implemented parallel readdirp with distribute xlator

Notes for users: Currently the directory listing gets slower as the number of bricks/nodes increases in a volume, though the file/directory numbers remain unchanged. With this feature, the performance of directory listing is made mostly independent of the number of nodes/bricks in the volume. Thus scale doesn’t exponentially reduce the directory listing performance. (On a 2, 5, 10, 25 brick setup we saw ~5, 100, 400, 450% improvement consecutively)

To enable this feature:

# gluster volume set <VOLNAME> performance.readdir-ahead on
# gluster volume set <VOLNAME> performance.parallel-readdir on

To disable this feature:

# gluster volume set <VOLNAME> performance.parallel-readdir off

If there are more than 50 bricks in the volume it is good to increase the cache size to be more than 10Mb (default value):

# gluster volume set <VOLNAME> performance.rda-cache-limit <CACHE SIZE>

Limitations:

Known Issues:

md-cache can optionally -ve cache security.ima xattr

Notes for users: From kernel version 3.X or greater, creating of a file results in removexattr call on security.ima xattr. This xattr is not set on the file unless IMA feature is active. With this patch, removxattr call returns ENODATA if it is not found in the cache.

The end benefit is faster create operations where IMA is not enabled.

To cache this xattr use,

# gluster volume set <VOLNAME> performance.cache-ima-xattrs on

The above option is on by default.

Limitations:

Known Issues:

Added support for CPU extensions in disperse computations

Notes for users: To improve disperse computations, a new way of generating dynamic code targeting specific CPU extensions like SSE and AVX on Intel processors is implemented. The available extensions are detected on run time. This can roughly double encoding and decoding speeds (or halve CPU usage).

This change is 100% compatible with the old method. No change is needed if an existing volume is upgraded.

You can control which extensions to use or disable them with the following command:

# gluster volume set <VOLNAME> disperse.cpu-extensions <type>

Valid values are:

The default value is ‘auto’. If a value is specified that is not detected on run-time, it will automatically fall back to the next available option.

Limitations:

Known Issues: To solve a conflict between the dynamic code generator and SELinux, it has been necessary to create a dynamic file on runtime in the directory /usr/libexec/glusterfs. This directory only exists if the server package is installed. On nodes with only the client package installed, this directory won’t exist and the dynamic code won’t be used.

It also needs root privileges to create the file there, so any gfapi application not running as root won’t be able to use dynamic code generation.

In these cases, disperse volumes will continue working normally but using the old implementation (equivalent to setting disperse.cpu-extensions to none).

More information and a discussion on how to solve this can be found here:

https://bugzilla.redhat.com/1421649

Bugs addressed

Bugs addressed since release-3.9 are listed below.

Quelle: https://gluster.readthedocs.io/en/latest/release-notes/3.10.0/

Die mobile Version verlassen