| title | Performance best practices for SQL Server on Linux |
|---|---|
| description | This article provides performance best practices and guidelines for running SQL Server on Linux. |
| author | tejasaks |
| ms.author | tejasaks |
| ms.reviewer | vanto, randolphwest |
| ms.date | 02/21/2023 |
| ms.service | sql |
| ms.subservice | linux |
| ms.topic | conceptual |
[!INCLUDE SQL Server - Linux]
This article provides best practices and recommendations to maximize performance for database applications that connect to [!INCLUDE ssnoversion-md] on Linux. These recommendations are specific to running on the Linux platform. All normal [!INCLUDE ssnoversion-md] recommendations, such as index design, still apply.
The following guidelines contain recommendations for configuring both [!INCLUDE ssnoversion-md] and the Linux Operating System (OS).
Consider using the following Linux OS configuration settings to experience the best performance for a [!INCLUDE ssnoversion-md] Installation.
The storage subsystem hosting data, transaction logs, and other associated files (such as checkpoint files for in-memory OLTP) should be capable of managing both average and peak workload gracefully. Normally, in on-premises environments, the storage vendor supports appropriate hardware RAID configuration with striping across multiple disks to ensure appropriate IOPS, throughput, and redundancy. Though, this can differ across different storage vendors and different storage offerings with varying architectures.
For [!INCLUDE ssnoversion-md] on Linux deployed on Azure Virtual Machines, consider using software RAID to ensure appropriate IOPS and throughput requirements are achieved. When configuring [!INCLUDE ssnoversion-md] on Azure virtual machines with similar storage considerations, see Storage configuration for SQL Server VMs.
The following example shows how to create software RAID in Linux on Azure Virtual Machines. Keep in mind that you should use the appropriate number of data disks for the required throughput and IOPS for volumes based on the data, transaction log, and tempdb I/O requirements. In the following example, eight data disks were attached to the Azure Virtual Machine; 4 to host data files, 2 for transaction logs, and 2 for tempdb workload.
# To locate the devices (for example /dev/sdc) for RAID creation, use the lsblk command
# For Data volume, using 4 devices, in RAID 5 configuration with 8KB stripes
mdadm --create --verbose /dev/md0 --level=raid5 --chunk=8K --raid-devices=4 /dev/sdc /dev/sdd /dev/sde /dev/sdf
# For Log volume, using 2 devices in RAID 10 configuration with 64KB stripes
mdadm --create --verbose /dev/md1 --level=raid10 --chunk=64K --raid-devices=2 /dev/sdg /dev/sdh
# For tempdb volume, using 2 devices in RAID 0 configuration with 64KB stripes
mdadm --create --verbose /dev/md2 --level=raid0 --chunk=64K --raid-devices=2 /dev/sdi /dev/sdjFor [!INCLUDE ssnoversion-md], you should use a RAID configuration. The deployed file system stripe unit (sunit) and stripe width should match the RAID geometry. For example, this is an XFS-based example for a log volume.
# Creating a log volume, using 6 devices, in RAID 10 configuration with 64KB stripes
mdadm --create --verbose /dev/md3 --level=raid10 --chunk=64K --raid-devices=6 /dev/sda /dev/sdb /dev/sdc /dev/sdd /dev/sde /dev/sdf
mkfs.xfs /dev/md3 -f -L log
meta-data=/dev/md3 isize=512 agcount=32, agsize=18287648 blks
= sectsz=4096 attr=2, projid32bit=1
= crc=1 finobt=1, sparse=1, rmapbt=0
= reflink=1
data = bsize=4096 blocks=585204384, imaxpct=5
= sunit=16 swidth=48 blks
naming =version 2 bsize=4096 ascii-ci=0, ftype=1
log =internal log bsize=4096 blocks=285744, version=2
= sectsz=4096 sunit=1 blks, lazy-count=1
realtime =none extsz=4096 blocks=0, rtextents=0The log array is a 6-drive RAID-10 with a 64-KB stripe. As you can see:
- For
sunit=16 blks, 16 * 4096 block size = 64 KB, matches the stripe size. - For
swidth=48 blks,swidth/sunit= 3, which is the number of data drives in the array, excluding parity drives.
[!INCLUDE ssnoversion-md] supports both EXT4 and XFS file systems to host the database, transaction logs, and additional files such as checkpoint files for in-memory OLTP in [!INCLUDE ssnoversion-md]. Microsoft recommends using XFS file system for hosting the [!INCLUDE ssnoversion-md] data and transaction log files.
# Formatting the volume with XFS filesystem
mkfs.xfs /dev/md0 -f -L datavolume
mkfs.xfs /dev/md1 -f -L logvolume
mkfs.xfs /dev/md2 -f -L tempdbNote
It's possible to configure the XFS file system to be case insensitive when creating and formatting the XFS volume. It is not the frequently used configuration in Linux ecosystem but can be used for compatibility reasons.
For example: mkfs.xfs /dev/md0 -f -n version=ci -L datavolume
In the example, parameters -n version=ci are used to configure the XFS filesystem to be case insensitive.
For the filesystem configuration on Persistent Memory devices, the block allocation for the underlying filesystem should be 2 MB. For more information on this article, review the article Technical considerations.
Your production environment may require more connections than the default open file limit of 1024. We recommend you set a soft limit of 16000, and a hard limit of 32727. For example, in RHEL, edit the /etc/security/limits.d/99-mssql-server.conf file to have the following values:
mssql hard nofile 32727
mssql soft nofile 16000To ensure that the drive(s) attached to the system remount automatically after a restart, add them to the /etc/fstab file. You should also use the UUID (Universally Unique Identifier) in /etc/fstab to refer to the drive, rather than just the device name (such as /dev/sdc1).
Use the noatime attribute with any file system that stores [!INCLUDE ssnoversion-md] data and log files. Refer to your Linux documentation on how to set this attribute. An example of how to enable noatime option for a volume mounted in Azure Virtual Machine follows.
The mount point entry in /etc/fstab:
UUID="xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx" /data1 xfs rw,attr2,noatime 0 0In the example above, UUID represents the device that you can find using the blkid command.
[!INCLUDE linux-forced-unit-access]
The following section describes the recommended Linux OS settings related to high performance and throughput for a [!INCLUDE ssnoversion-md] installation. See your Linux distribution's documentation for the process to configure these settings. You can use TuneD as described, to configure many CPUs and kernel configurations, described in the next section.
For Red Hat Enterprise Linux (RHEL) users, the TuneD throughput-performance profile configures some kernel and CPU settings automatically (except for C-States). Starting with RHEL 8.0, a TuneD profile named mssql was codeveloped with Red Hat and offers finer Linux performance-related tunings for [!INCLUDE ssnoversion-md] workloads. This profile includes the RHEL throughput-performance profile, and we present its definitions below for your review with other Linux distributions and RHEL releases without this profile.
For SUSE Linux Enterprise Server 12 SP5, Ubuntu 18.04, and Red Hat Enterprise Linux 7.x, the tuned package can be installed manually. It can be used to create and configure the mssql profile as described below.
#
# A TuneD configuration for SQL Server on Linux
#
[main]
summary=Optimize for Microsoft SQL Server
include=throughput-performance
[cpu]
force_latency=5
[sysctl]
vm.swappiness = 1
vm.dirty_background_ratio = 3
vm.dirty_ratio = 80
vm.dirty_expire_centisecs = 500
vm.dirty_writeback_centisecs = 100
vm.transparent_hugepages=always
# For multi-instance SQL deployments, use
# vm.transparent_hugepages=madvise
vm.max_map_count=1600000
net.core.rmem_default = 262144
net.core.rmem_max = 4194304
net.core.wmem_default = 262144
net.core.wmem_max = 1048576
kernel.numa_balancing=0
#Note: If you are using Linux distributions with kernel versions greater than 4.18, please comment the following options as shown; otherwise, uncomment the following options if you are using distributions with kernel versions less than 4.18.
# kernel.sched_latency_ns = 60000000
# kernel.sched_migration_cost_ns = 500000
# kernel.sched_min_granularity_ns = 15000000
# kernel.sched_wakeup_granularity_ns = 2000000To enable this TuneD profile, save these definitions in a tuned.conf file under a /usr/lib/tuned/mssql folder, and enable the profile using the following commands:
chmod +x /usr/lib/tuned/mssql/tuned.conf
tuned-adm profile mssqlVerify that the profile is active, with the following command:
tuned-adm activeor
tuned-adm listThe following table provides recommendations for CPU settings:
| Setting | Value | More information |
|---|---|---|
| CPU frequency governor | performance | See the cpupower command |
| ENERGY_PERF_BIAS | performance | See the x86_energy_perf_policy command |
| min_perf_pct | 100 | See your documentation on intel p-state |
| C-States | C1 only | See your Linux or system documentation on how to ensure C-States is set to C1 only |
Using TuneD as described earlier automatically configures CPU frequency governor, ENERGY_PERF_BIAS, and min_perf_pct settings appropriately due to the throughput-performance profile being used as base for the mssql profile. C-States parameter must be configured manually according to the documentation provided by Linux or the system distributor.
The following table provides recommendations for disk settings:
| Setting | Value | More information |
|---|---|---|
Disk readahead |
4096 | See the blockdev command |
| sysctl settings | kernel.sched_min_granularity_ns = 15000000 kernel.sched_wakeup_granularity_ns = 2000000 vm.dirty_ratio = 80 vm.dirty_background_ratio = 3 vm.swappiness = 1 |
See the sysctl command |
Description:
-
vm.swappiness: This parameter controls relative weight given to swapping out runtime process memory as compared to filesystem cache. The default value for this parameter is 60, which indicates swapping runtime process memory pages as compared to removing filesystem cache pages at ratio of 60:140. Setting the value 1 indicates strong preference for keeping runtime process memory in physical memory at expense of filesystem cache. Since [!INCLUDE ssnoversion-md] uses buffer pool as a data page cache and strongly prefers to write through to physical hardware bypassing filesystem cache for reliable recovery, aggressive swappiness configuration can be beneficial for high performing and dedicated [!INCLUDE ssnoversion-md]. You can find additional information at Documentation for /proc/sys/vm/ - #swappiness -
vm.dirty_*: [!INCLUDE ssnoversion-md] file write accesses are uncached, satisfying its data integrity requirements. These parameters allow efficient asynchronous write performance and lower the storage IO impact of Linux caching writes by allowing large enough caching while throttling flushing. -
kernel.sched_*: These parameter values represent the current recommendation for tweaking the Completely Fair Scheduling (CFS) algorithm in the Linux Kernel, to improve throughput of network and storage IO calls with respect to inter-process preemption and resumption of threads.
Using the mssql TuneD profile configures the vm.swappiness, vm.dirty_* and kernel.sched_* settings. The disk readahead configuration using blockdev command is per device and must be performed manually.
If you install [!INCLUDE ssnoversion-md] on a multi-node NUMA system, the following kernel.numa_balancing kernel setting is enabled by default. To allow [!INCLUDE ssnoversion-md] to operate at maximum efficiency on a NUMA system, disable auto NUMA balancing on a multi-node NUMA system:
sysctl -w kernel.numa_balancing=0Using the mssql TuneD profile configures the kernel.numa_balancing option.
The default setting of vm.max_map_count (which is 65536) may not be high enough for a [!INCLUDE ssnoversion-md] installation. For this reason, change the vm.max_map_count value to at least 262144 for a [!INCLUDE ssnoversion-md] deployment, and refer to the Proposed Linux settings using a TuneD mssql profile section for further tunings of these kernel parameters. The max value for vm.max_map_count is 2147483647.
sysctl -w vm.max_map_count=1600000Using the mssql TuneD profile configures the vm.max_map_count option.
Most Linux installations should have this option on by default. We recommend for the most consistent performance experience to leave this configuration option enabled. However, if there is high memory paging activity in [!INCLUDE ssnoversion-md] deployments with multiple instances, for example, or [!INCLUDE ssnoversion-md] execution with other memory demanding applications on the server, we suggest testing your applications performance after executing the following command:
echo madvise > /sys/kernel/mm/transparent_hugepage/enabledOr modify the mssql TuneD profile with the line:
vm.transparent_hugepages=madviseAnd make the mssql profile is active after the modification:
tuned-adm off
tuned-adm profile mssqlUsing the mssql TuneD profile configures the transparent_hugepage option.
Like there are storage and CPU recommendations, there are Network specific recommendations as well listed below for reference. Not all settings in the following examples are available across different NICs. Refer and consult with NIC vendors for guidance for each of these options. Test and configure this on development environments before applying them on production environments. The following options are explained with examples, and the commands used are specific to NIC type and vendor.
-
Configuring network port buffer size: In the example below, the NIC is named 'eth0', which is an Intel-based NIC. For Intel based NIC, the recommended buffer size is 4 KB (4096). Verify the pre-set maximums and then configure it using the following example:
#To check the pre-set maximums please run the command, example NIC name used here is:"eth0" ethtool -g eth0 #command to set both the rx (receive) and tx (transmit) buffer size to 4 KB. ethtool -G eth0 rx 4096 tx 4096 #command to check the value is properly configured is: ethtool -g eth0
-
Enable jumbo frames: Before enabling jumbo frames, verify that all the network switches, routers, and anything else essential in the network packet path between the clients and the [!INCLUDE ssnoversion-md] support jumbo frames. Only then, enabling jumbo frames can improve performance. After jumbo frames are enabled, connect to [!INCLUDE ssnoversion-md] and change the network packet size to 8060 using
sp_configureas shown below:#command to set jumbo frame to 9014 for a Intel NIC named eth0 is ifconfig eth0 mtu 9014 #verify the setting using the command: ip addr | grep 9014
EXEC sp_configure 'network packet size', '8060'; GO RECONFIGURE WITH OVERRIDE; GO
-
By default, we recommend setting the port for adaptive RX/TX IRQ coalescing, meaning interrupt delivery is adjusted to improve latency when packet rate is low and improve throughput when packet rate is high. This setting might not be available across all the different network infrastructure, so review the existing network infrastructure and confirm that this is supported. The example below is for the NIC named 'eth0', which is an intel-based NIC:
#command to set the port for adaptive RX/TX IRQ coalescing ethtool -C eth0 adaptive-rx on ethtool -C eth0 adaptive-tx on #confirm the setting using the command: ethtool -c eth0
[!NOTE]
For a predictable behavior for high-performance environments, like environments for benchmarking, disable the adaptive RX/TX IRQ coalescing and then set specifically the RX/TX interrupt coalescing. See the example commands to disable the RX/TX IRQ coalescing and then specifically set the values:#commands to disable adaptive RX/TX IRQ coalescing ethtool -C eth0 adaptive-rx off ethtool -C eth0 adaptive-tx off #confirm the setting using the command: ethtool -c eth0 #Let us set the rx-usecs parameter which specify how many microseconds after at least 1 packet is received before generating an interrupt, and the [irq] parameters are the corresponding delays in updating the #status when the interrupt is disabled. For Intel bases NICs below are good values to start with: ethtool -C eth0 rx-usecs 100 tx-frames-irq 512 #confirm the setting using the command: ethtool -c eth0
-
We also recommend RSS (Receive-Side Scaling) enabled and by default, combining the RX and TX side of RSS queues. There have been specific scenarios, when working with Microsoft Support, where disabling RSS has improved the performance as well. Test this setting in test environments before applying it on production environments. The following example is for Intel NICs.
#command to get pre-set maximums ethtool -l eth0 #note the pre-set "Combined" maximum value. let's consider for this example, it is 8. #command to combine the queues with the value reported in the pre-set "Combined" maximum value: ethtool -L eth0 combined 8 #you can verify the setting using the command below ethtool -l eth0
-
Working with NIC port IRQ affinity. To achieve expected performance by tweaking the IRQ affinity, consider few important parameters like Linux handling of the server topology, NIC driver stack, default settings, and irqbalance setting. Optimizations of the NIC port IRQ affinities settings are done with the knowledge of server topology, disabling the irqbalance, and using the NIC vendor-specific settings.
The following example of Mellanox specific network infrastructure helps to explain the configuration. For more information, see Performance Tuning tools for Mellanox Network Adapters. The commands change based on the environment. Contact the NIC vendor for further guidance:
#disable irqbalance or get a snapshot of the IRQ settings and force the daemon to exit systemctl disable irqbalance.service #or irqbalance --oneshot #download the Mellanox mlnx tools -- see https://support.mellanox.com/s/article/MLNX2-117-2523kn #be sure, common_irq_affinity.sh is executable. if not, #chmod +x common_irq_affinity.sh #display IRQ affinity for Mellanox NIC port; e.g eth0 ./show_irq_affinity.sh eth0 #optimize for best throughput performance with a Mellanox tool ./mlnx_tune -p HIGH_THROUGHPUT #set hardware affinity to the NUMA node hosting physically the NIC and its port ./set_irq_affinity_bynode.sh `\cat /sys/class/net/eth0/device/numa_node` eth0 #verify IRQ affinity ./show_irq_affinity.sh eth0 #add IRQ coalescing optimizations ethtool -C eth0 adaptive-rx off ethtool -C eth0 adaptive-tx off ethtool -C eth0 rx-usecs 750 tx-frames-irq 2048 #verify the settings ethtool -c eth0
-
After the above changes are done, verify the speed of the NIC to ensure it matches the expectation using the following command:
ethtool eth0 | grep -i Speed
-
For best storage IO performance, use Linux multiqueue scheduling for block devices, which enables the block layer performance to scale well with fast solid-state drives (SSDs) and multi-core systems. Check the documentation if it is enabled by default in your Linux distribution. In most other cases, booting the kernel with
scsi_mod.use_blk_mq=yenables it, though documentation of the Linux distribution in use may have further guidance on it. This is consistent with the upstream Linux kernel. -
As multipath IO is often used for [!INCLUDE ssnoversion-md] deployments, configure the device mapper (DM) multi-queue target to use the
blk-mqinfrastructure, by enabling thedm_mod.use_blk_mq=ykernel boot option. The default value isn(disabled). This setting, when the underlying SCSI devices are usingblk-mq, reduces locking overhead at the DM layer. For more information on how to configure multipath IO, refer to your Linux distribution's documentation.
Ensure you have a properly configured swapfile to avoid any out of memory issues. Consult your Linux documentation for how to create and properly size a swapfile.
If you're running [!INCLUDE ssnoversion-md] on Linux in a virtual machine, make sure you select options to fix the amount of memory reserved for the virtual machine. Don't use features like Hyper-V Dynamic Memory.
Perform the following configuration tasks after you install [!INCLUDE ssnoversion-md] on Linux to achieve best performance for your application.
-
Use PROCESS AFFINITY for node and/or CPUs
Use
ALTER SERVER CONFIGURATIONto setPROCESS AFFINITYfor all theNUMANODEs and/or CPUs you're using for [!INCLUDE ssnoversion-md] (which is typically for all NODEs and CPUs) on a Linux OS. Processor affinity helps maintain efficient Linux and SQL Scheduling behavior. Using theNUMANODEoption is the simplest method. UsePROCESS AFFINITYeven if you have only a single NUMA Node on your computer. For more information on how to setPROCESS AFFINITY, see the ALTER SERVER CONFIGURATION article. -
Configure multiple
tempdbdata filesBecause a [!INCLUDE ssnoversion-md] on Linux installation doesn't offer an option to configure multiple
tempdbfiles, we recommend that you consider creating multipletempdbdata files after installation. For more information, see the guidance in the article, Recommendations to reduce allocation contention in SQL Server tempdb database.
The following recommendations are optional configuration settings that you may choose to perform after installation of [!INCLUDE ssnoversion-md] on Linux. These choices are based on the requirements of your workload and configuration of your Linux OS.
-
Set a memory limit with mssql-conf
In order to ensure There's enough free physical memory for the Linux OS, the [!INCLUDE ssnoversion-md] process uses only 80% of the physical RAM by default. For some systems with large amount of physical RAM, 20% might be a significant number. For example, on a system with 1 TB of RAM, the default setting would leave around 200 GB of RAM unused. In this situation, you might want to configure the memory limit to a higher value. See the documentation on the mssql-conf tool and the memory.memorylimitmb setting that controls the memory visible to [!INCLUDE ssnoversion-md] (in units of MB).
When changing this setting, be careful not to set this value too high. If you don't leave enough memory, you could experience problems with the Linux OS and other Linux applications.