pvcreate --dataalignment alignment_value disk
Configuring GlusterFS for Enhancing Performance
This chapter provides information on configuring GlusterFS
and explains clear and simple activities that can improve system
performance. A script that encodes the best-practice recommendations in
this section is located at
/usr/lib/glusterfs/.unsupported/rhs-system-init.sh. You can refer the
same for more information.
GlusterFS includes support for JBOD (Just a Bunch of Disks). In the JBOD configuration, a single physical disk serves as storage for a GlusterFS brick. JBOD is supported with three-way replication. GlusterFS in JBOD configuration is recommended for highly multi-threaded workloads with sequential reads to large files. For such workloads, JBOD results in more efficient use of disk bandwidth by reducing disk head movement from concurrent accesses. For other workloads, two-way replication with hardware RAID is recommended.
The RAID levels that are most commonly recommended are RAID 6 and RAID 10. RAID 6 provides better space efficiency, good read performance and good performance for sequential writes to large files.
When configured across 12 disks, RAID 6 can provide ~40% more storage space in comparison to RAID 10, which has a 50% reduction in capacity. However, RAID 6 performance for small file writes and random writes tends to be lower than RAID 10. If the workload is strictly small files, then RAID 10 is the optimal configuration.
An important parameter in hardware RAID configuration is the stripe unit size. With thin provisioned disks, the choice of RAID stripe unit size is closely related to the choice of thin-provisioning chunk size.
For RAID 10, a stripe unit size of 256 KiB is recommended.
For RAID 6, the stripe unit size must be chosen such that the full stripe size (stripe unit * number of data disks) is between 1 MiB and 2 MiB, preferably in the lower end of the range. Hardware RAID controllers usually allow stripe unit sizes that are a power of 2. For RAID 6 with 12 disks (10 data disks), the recommended stripe unit size is 128KiB.
Support for JBOD has the following limitations:
Each server in the JBOD configuration can have a maximum of 24 disks.
Three-way replication must be used when using JBOD.
In the JBOD configuration, physical disks are not aggregated into RAID devices, but are visible as separate disks to the operating system. This simplifies system configuration by not requiring a hardware RAID controller.
If disks on the system are connected through a hardware RAID controller,
refer to the RAID controller documentation on how to create a JBOD
configuration; typically, JBOD is realized by exposing
raw drives to
the operating system using a
Format bricks using the following configurations to enhance performance:
The steps for creating a brick from a physical device is listed below. An outline of steps for creating multiple bricks on a physical device is listed as Example - Creating multiple bricks on a physical device below.
Creating the Physical Volume
pvcreatecommand is used to create the physical volume. The Logical Volume Manager can use a portion of the physical volume for storing its metadata while the rest is used as the data portion.Align the I/O at the Logical Volume Manager (LVM) layer using
--dataalignmentoption while creating the physical volume.
The command is used in the following format:
For JBOD, use an alignment value of
In case of hardware RAID, the alignment_value should be obtained by multiplying the RAID stripe unit size with the number of data disks. If 12 disks are used in a RAID 6 configuration, the number of data disks is 10; on the other hand, if 12 disks are used in a RAID 10 configuration, the number of data disks is 6.
For example, the following command is appropriate for 12 disks in a RAID 6 configuration with a stripe unit size of 128 KiB:
# pvcreate --dataalignment 1280k disk
The following command is appropriate for 12 disks in a RAID 10 configuration with a stripe unit size of 256 KiB:
# pvcreate --dataalignment 1536k disk
To view the previously configured physical volume settings for
--dataalignment, run the following command:
# pvs -o +pe_start disk PV VG Fmt Attr PSize PFree 1st PE /dev/sdb lvm2 a-- 9.09t 9.09t 1.25m
Creating the Volume Group
The volume group is created using the
For hardware RAID, in order to ensure that logical volumes created in the volume group are aligned with the underlying RAID geometry, it is important to use the
-- physicalextentsizeoption. Execute the
vgcreatecommand in the following format:
# vgcreate --physicalextentsize extent_size VOLGROUP physical_volume
The extent_size should be obtained by multiplying the RAID stripe unit size with the number of data disks. If 12 disks are used in a RAID 6 configuration, the number of data disks is 10; on the other hand, if 12 disks are used in a RAID 10 configuration, the number of data disks is 6.
For example, run the following command for RAID-6 storage with a stripe unit size of 128 KB, and 12 disks (10 data disks):
# vgcreate --physicalextentsize 1280k VOLGROUP physical_volume
In the case of JBOD, use the
vgcreatecommand in the following format:
# vgcreate VOLGROUP physical_volume
Creating the Thin Pool
A thin pool provides a common pool of storage for thin logical volumes (LVs) and their snapshot volumes, if any.
Execute the following command to create a thin-pool:
# lvcreate --thinpool VOLGROUP/thin_pool --size <pool_size> --chunksize <chunk_size> --poolmetadatasize <meta_size> --zero n
Internally, a thin pool contains a separate metadata device that is used to track the (dynamically) allocated regions of the thin LVs and snapshots. The
poolmetadatasizeoption in the above command refers to the size of the pool meta data device. + The maximum possible size for a metadata LV is 16 GiB. GlusterFS recommends creating the metadata device of the maximum supported size. You can allocate less than the maximum if space is a concern, but in this case you should allocate a minimum of 0.5% of the pool size.
An important parameter to be specified while creating a thin pool is the chunk size,which is the unit of allocation. For good performance, the chunk size for the thin pool and the parameters of the underlying hardware RAID storage should be chosen so that they work well together. + For RAID-6 storage, the striping parameters should be chosen so that the full stripe size (stripe_unit size * number of data disks) is between 1 MiB and 2 MiB, preferably in the low end of the range. The thin pool chunk size should be chosen to match the RAID 6 full stripe size. Matching the chunk size to the full stripe size aligns thin pool allocations with RAID 6 stripes, which can lead to better performance. Limiting the chunk size to below 2 MiB helps reduce performance problems due to excessive copy-on-write when snapshots are used. + For example, for RAID 6 with 12 disks (10 data disks), stripe unit size should be chosen as 128 KiB. This leads to a full stripe size of 1280 KiB (1.25 MiB). The thin pool should then be created with the chunk size of 1280 KiB. + For RAID 10 storage, the preferred stripe unit size is 256 KiB. This can also serve as the thin pool chunk size. Note that RAID 10 is recommended when the workload has a large proportion of small file writes or random writes. In this case, a small thin pool chunk size is more appropriate, as it reduces copy-on-write overhead with snapshots. + For JBOD, use a thin pool chunk size of 256 KiB.
- block zeroing
By default, the newly provisioned chunks in a thin pool are zeroed to prevent data leaking between different block devices. In the case of GlusterFS, where data is accessed via a file system, this option can be turned off for better performance with the
--zero noption. Note that
ndoes not need to be replaced. + The following example shows how to create the thin pool: +
lvcreate --thinpool VOLGROUP/thin_pool --size 800g --chunksize 1280k --poolmetadatasize 16G --zero n
Creating a Thin Logical Volume
After the thin pool has been created as mentioned above, a thinly provisioned logical volume can be created in the thin pool to serve as storage for a brick of a GlusterFS volume.
# lvcreate --thin --name LV_name --virtualsize LV_size VOLGROUP/thin_pool
Example - Creating multiple bricks on a physical device
The steps above (LVM Layer) cover the case where a single brick is being created on a physical device. This example shows how to adapt these steps when multiple bricks need to be created on a physical device.
In this following steps, we are assuming the following:
Two bricks must be created on the same physical device
One brick must be of size 4 TiB and the other is 2 TiB
The device is
/dev/sdb, and is a RAID-6 device with 12 disks
The 12-disk RAID-6 device has been created according to the recommendations in this chapter, that is, with a stripe unit size of 128 KiB
Create a single physical volume using pvcreate
# pvcreate --dataalignment 1280k /dev/sdb
Create a single volume group on the device
# vgcreate --physicalextentsize 1280k vg1 /dev/sdb
Create a separate thin pool for each brick using the following commands:
# lvcreate --thinpool vg1/thin_pool_1 --size 4T --chunksize 1280K --poolmetadatasize 16G --zero n
# lvcreate --thinpool vg1/thin_pool_2 --size 2T --chunksize 1280K --poolmetadatasize 16G --zero n
In the examples above, the size of each thin pool is chosen to be the same as the size of the brick that will be created in it. With thin provisioning, there are many possible ways of managing space, and these options are not discussed in this chapter.
Create a thin logical volume for each brick
# lvcreate --thin --name lv1 --virtualsize 4T vg1/thin_pool_1
# lvcreate --thin --name lv2 --virtualsize 2T vg1/thin_pool_2
Follow the XFS Recommendations (next step) in this chapter for creating and mounting filesystems for each of the thin logical volumes
mkfs.xfs options /dev/vg1/lv1
mkfs.xfs options /dev/vg1/lv2
mount options /dev/vg1/lv1 mount_point_1
mount options /dev/vg1/lv2 mount_point_2
XFS Inode Size
As GlusterFS makes extensive use of extended attributes, an XFS inode size of 512 bytes works better with GlusterFS than the default XFS inode size of 256 bytes. So, inode size for XFS must be set to 512 bytes while formatting the GlusterFS bricks. To set the inode size, you have to use -i size option with the
mkfs.xfscommand as shown in the following Logical Block Size for the Directory section.
XFS RAID Alignment
When creating an XFS file system, you can explicitly specify the striping parameters of the underlying storage in the following format:
mkfs.xfs other_options -d su=stripe_unit_size,sw=stripe_width_in_number_of_disks device
For RAID 6, ensure that I/O is aligned at the file system layer by providing the striping parameters. For RAID 6 storage with 12 disks, if the recommendations above have been followed, the values must be as following:
# mkfs.xfs other_options -d su=128k,sw=10 device
For RAID 10 and JBOD, the
-d su=<>,sw=<>option can be omitted. By default, XFS will use the thin-p chunk size and other parameters to make layout decisions.
Logical Block Size for the Directory
An XFS file system allows to select a logical block size for the file system directory that is greater than the logical block size of the file system. Increasing the logical block size for the directories from the default 4 K, decreases the directory I/O, which in turn improves the performance of directory operations. To set the block size, you need to use
-n sizeoption with the
mkfs.xfscommand as shown in the following example output.
Following is the example output of RAID 6 configuration along with inode and block size options:
# mkfs.xfs -f -i size=512 -n size=8192 -d su=128k,sw=10 logical volume meta-data=/dev/mapper/gluster-brick1 isize=512 agcount=32, agsize=37748736 blks = sectsz=512 attr=2, projid32bit=0 data = bsize=4096 blocks=1207959552, imaxpct=5 = sunit=32 swidth=320 blks naming = version 2 bsize=8192 ascii-ci=0 log =internal log bsize=4096 blocks=521728, version=2 = sectsz=512 sunit=32 blks, lazy-count=1 realtime =none extsz=4096 blocks=0, rtextents=0
inode32 and inode64 are two most common allocation strategies for XFS. With inode32 allocation strategy, XFS places all the inodes in the first 1 TiB of disk. With larger disk, all the inodes would be stuck in first 1 TiB. inode32 allocation strategy is used by default.
With inode64 mount option inodes would be replaced near to the data which would be minimize the disk seeks.
To set the allocation strategy to inode64 when file system is being mounted, you need to use
-o inode64 `option with the `mountcommand as shown in the following Access Time section.
If the application does not require to update the access time on files, than file system must always be mounted with
noatimemount option. For example:
# mount -t xfs -o inode64,noatime <logical volume> <mount point>
This optimization improves performance of small-file reads by avoiding updates to the XFS inodes when files are read.
/etc/fstab entry for option E + F <logical volume> <mount point>xfs inode64,noatime 0 0
Each XFS file system is partitioned into regions called allocation groups. Allocation groups are similar to the block groups in ext3, but allocation groups are much larger than block groups and are used for scalability and parallelism rather than disk locality. The default allocation for an allocation group is 1 TiB.
Allocation group count must be large enough to sustain the concurrent allocation workload. In most of the cases allocation group count chosen by
mkfs.xfscommand would give the optimal performance. Do not change the allocation group count chosen by
mkfs.xfs, while formatting the file system.
Percentage of space allocation to inodes
If the workload is very small files (average file size is less than 10 KB ), then it is recommended to set
10, while formatting the file system.
For small-file and random write performance, we strongly recommend writeback cache, that is, non-volatile random-access memory (NVRAM) in your storage controller. For example, normal Dell and HP storage controllers have it. Ensure that NVRAM is enabled, that is, the battery is working. Refer your hardware documentation for details on enabling NVRAM.
Do not enable writeback caching in the disk drives, this is a policy where the disk drive considers the write is complete before the write actually made it to the magnetic media (platter). As a result, the disk write cache might lose its data during a power failure or even loss of metadata leading to file system corruption.
Data traffic Network becomes a bottleneck as and when number of storage nodes increase. By adding a 10GbE or faster network for data traffic, you can achieve faster per node performance. Jumbo frames must be enabled at all levels, that is, client , GlusterFS node, and ethernet switch levels. MTU of size N+208 must be supported by ethernet switch where N=9000. We recommend you to have a separate network for management and data traffic when protocols like NFS /CIFS are used instead of native client. Preferred bonding mode for GlusterFS client is mode 6 (balance-alb), this allows client to transmit writes in parallel on separate NICs much of the time.
GlusterFS does not consume significant compute resources from the storage nodes themselves. However, read intensive workloads can benefit greatly from additional RAM.
Virtual Memory Parameters
The data written by the applications is aggregated in the operating system page cache before being flushed to the disk. The aggregation and writeback of dirty data is governed by the Virtual Memory parameters. The following parameters may have a significant performance impact:
The appropriate values of these parameters vary with the type of workload:
Large-file sequential I/O workloads benefit from higher values for these parameters.
For small-file and random I/O workloads it is recommended to keep these parameter values low.
The GlusterFS tuned profiles set the values for these parameters appropriately. Hence, it is important to select and activate the appropriate GlusterFS profile based on the workload.
Small File Performance Enhancements
The ratio of the time taken to perform operations on the metadata of a
file to performing operations on its data determines the difference
between large files and small files.
Metadata-intensive workload is
the term used to identify such workloads. A few performance enhancements
can be made to optimize the network and storage performance and minimize
the effect of slow throughput and response time for small files in a
GlusterFS trusted storage pool.
For a small-file workload, activate the
Configuring Threads for Event Processing.
You can set the
for the client and server components. Setting the value to 3, for
example, would enable handling three network connections simultaneously.
Setting the event threads value for a client
You can tune the GlusterFS Server performance by tuning the event thread values.
# gluster volume set VOLNAME client.event-threads <value>
# gluster volume set test-vol client.event-threads 3
Setting the event thread value for a server
You can tune the GlusterFS Server performance using event thread values.
# gluster volume set VOLNAME server.event-threads <value>
# gluster volume set test-vol server.event-threads 3
Verifying the event thread values
You can verify the event thread values that are set for the client and server components by executing the following command:
# gluster volume info VOLNAME
See topic, Configuring Volume Options for information on the minimum, maximum, and default values for setting these volume options.
Best practices to tune event threads.
It is possible to see performance gains with the GlusterFS stack by tuning the number of threads processing events from network connections.The following are the recommended best practices to tune the event thread values.
As each thread processes a connection at a time, having more threads than connections to either the brick processes (
glusterfsd) or the client processes (
gfapi) is not recommended. Due to this reason, monitor the connection counts (using the
netstatcommand) on the clients and on the bricks to arrive at an appropriate number for the event thread count.
Configuring a higher event threads value than the available processing units could again cause context switches on these threads. As a result reducing the number deduced from the previous step to a number that is less that the available processing units is recommended.
If a GlusterFS volume has a high number of brick processes running on a single node, then reducing the event threads number deduced in the previous step would help the competing processes to gain enough concurrency and avoid context switches across the threads.
If a specific thread consumes more number of CPU cycles than needed, increasing the event thread count would enhance the performance of the GlusterFS Server.
In addition to the deducing the appropriate event-thread count, increasing the `server.outstanding-rpc-limit `on the storage nodes can also help to queue the requests for the brick processes and not let the requests idle on the network queue.
Another parameter that could improve the performance when tuning the event-threads value is to set the` performance.io-thread-count` (and its related thread-counts) to higher values, as these threads perform the actual IO operations on the underlying file system.
Enabling Lookup Optimization
Distribute xlator (DHT) has a performance penalty when it deals with negative lookups. Negative lookups are lookup operations for entries that does not exist in the volume. A lookup for a file/directory that does not exist is a negative lookup.
Negative lookups are expensive and typically slows down file creation, as DHT attempts to find the file in all sub-volumes. This especially impacts small file performance, where a large number of files are being added/created in quick succession to the volume.
The negative lookup fan-out behavior can be optimized by not performing the same in a balanced volume.
cluster.lookup-optimize configuration option enables DHT lookup
optimization. To enable this option run the following command:
# gluster volume set VOLNAME cluster.lookup-optimize <on/off>\
The configuration takes effect for newly created directories immediately post setting the above option. For existing directories, a rebalance is required to ensure the volume is in balance before DHT applies the optimization on older directories.
If a system is configured for two ways, active-active replication, write throughput will generally be half of what it would be in a non-replicated configuration. However, read throughput is generally improved by replication, as reads can be delivered from either storage node.