Doc Menu

Operating System Tuning

This document introduces how to tune each subsystem of CentOS 7.

Note:

  • The default configuration of the CentOS 7 operating system is suitable for most services running under moderate workloads. Adjusting the performance of a particular subsystem might negatively affects other subsystems. Therefore, before tuning the system, back up all the user data and configuration information.
  • Fully test all the changes in the test environment before applying them to the production environment.

Performance analysis methods

System tuning must be based on the results of system performance analysis. This section lists common methods for performance analysis.

In 60 seconds

Linux Performance Analysis in 60,000 Milliseconds is published by the author Brendan Gregg and the Netflix Performance Engineering team. All tools used can be obtained from the official release of Linux. You can analyze outputs of the following list items to troubleshoot most common performance issues.

  • uptime
  • dmesg | tail
  • vmstat 1
  • mpstat -P ALL 1
  • pidstat 1
  • iostat -xz 1
  • free -m
  • sar -n DEV 1
  • sar -n TCP,ETCP 1
  • top

For detailed usage, see the corresponding man instructions.

perf

perf is an important performance analysis tool provided by the Linux kernel, which covers hardware level (CPU/PMU, performance monitoring unit) features and software features (software counters, trace points). For detailed usage, see perf Examples.

BCC/bpftrace

Starting from CentOS 7.6, the Linux kernel has supported Berkeley Packet Filter (BPF). Therefore, you can choose proper tools to conduct an in-depth analysis based on the results in In 60 seconds. Compared with perf/ftrace, BPF provides programmability and smaller performance overhead. Compared with kprobe, BPF provides higher security and is more suitable for the production environments. For detailed usage of the BCC toolkit, see BPF Compiler Collection (BCC).

Performance tuning

This section introduces performance tuning based on the classified kernel subsystems.

CPU—frequency scaling

cpufreq is a module that dynamically adjusts the CPU frequency. It supports five modes. To ensure service performance, select the performance mode and fix the CPU frequency at the highest supported operating frequency without dynamic adjustment. The command for this operation is cpupower frequency-set --governor performance.

CPU—interrupt affinity

  • Automatic balance can be implemented through the irqbalance service.
  • Manual balance:
    • Identify the devices that need to balance interrupts. Starting from CentOS 7.5, the system automatically configures the best interrupt affinity for certain devices and their drivers, such as devices that use the be2iscsi driver and NVMe settings. You can no longer manually configure interrupt affinity for such devices.
    • For other devices, check the chip manual to see whether these devices support distributing interrupts.
      • If they do not, all interrupts of these devices are routed to the same CPU and cannot be modified.
      • If they do, calculate the smp_affinity mask and set the corresponding configuration file. For details, see the kernel document.

NUMA CPU binding

To avoid accessing memory across Non-Uniform Memory Access (NUMA) nodes as much as possible, you can bind a thread/process to certain CPU cores by setting the CPU affinity of the thread. For ordinary programs, you can use the numactl command for the CPU binding. For detailed usage, see the Linux manual pages. For network interface card (NIC) interrupts, see tune network.

Memory—transparent huge page (THP)

It is NOT recommended to use THP for database applications, because databases often have sparse rather than continuous memory access patterns. If high-level memory fragmentation is serious, a higher latency will occur when THP pages are allocated. If the direct compaction is enabled for THP, the CPU usage will surge. Therefore, it is recommended to disable THP.

echo never > /sys/kernel/mm/transparent_hugepage/enabled
echo never > /sys/kernel/mm/transparent_hugepage/defrag

Memory—virtual memory parameters

  • dirty_ratio percentage ratio. When the total amount of dirty page caches reach this percentage ratio of the total system memory, the system starts to use the pdflush operation to write the dirty page caches to disk. The default value of dirty_ratio is 20% and usually does not need adjustment. For high-performance SSDs such as NVMe devices, lowering this value helps improve the efficiency of memory reclamation.
  • dirty_background_ratio percentage ratio. When the total amount of dirty page caches reach this percentage ratio of the total system memory, the system starts to write the dirty page caches to the disk in the background. The default value of dirty_ratio is 10% and usually does not need adjustment. For high-performance SSDs such as NVMe devices, setting a lower value helps improve the efficiency of memory reclamation.

Storage and file system

The core I/O stack link is long, including the file system layer, the block device layer, and the driver layer.

I/O scheduler

The I/O scheduler determines when and how long I/O operations run on the storage device. It is also called I/O elevator. For SSD devices, it is recommended to set the I/O scheduling policy to noop.

echo noop > /sys/block/${SSD_DEV_NAME}/queue/scheduler

Formatting parameters—block size

Blocks are the working units of the file system. The block size determines how much data can be stored in a single block, and thus determines the minimum amount of data to be written or read each time.

The default block size is suitable for most scenarios. However, if the block size (or the size of multiple blocks) is the same or slightly larger than the amount of data normally read or written each time, the file system performs better and the data storage efficiency is higher. Small files still use the entire block. Files can be distributed among multiple blocks, but this will increase runtime overhead.

When using the mkfs command to format a device, specify the block size as a part of the file system options. The parameters that specify the block size vary with the file system. For details, see the corresponding mkfs manual pages, such as using man mkfs.ext4.

mount parameters

If the noatime option is enabled in the mount command, the update of metadata is disabled when files are read. If the nodiratime behavior is enabled, the update of metadata is disabled when the directory is read.

Network tuning

The network subsystem consists of many different parts with sensitive connections. The CentOS 7 network subsystem is designed to provide the best performance for most workloads and automatically optimizes the performance of these workloads. Therefore, usually you do not need to manually adjust network performance.

Network issues are usually caused by issues of hardware or related devices. So before tuning the protocol stack, rule out hardware issues.

Although the network stack is largely self-optimizing, the following aspects in the network packet processing might become the bottleneck and affect performance:

  • NIC hardware cache: To correctly observe the packet loss at the hardware level, use the ethtool -S ${NIC_DEV_NAME} command to observe the drops field. When packet loss occurs, it might be that the processing speed of the hard/soft interrupts cannot catch up with the receiving speed of NIC. If the received buffer size is less than the upper limit, you can also try to increase the RX buffer to avoid packet loss. The query command is: ethtool -g ${NIC_DEV_NAME}, and the modification command is ethtool -G ${NIC_DEV_NAME}.

  • Hardware interrupts: If the NIC supports the Receive-Side Scaling (RSS, also called multi-NIC receiving) feature, observe the /proc/interrupts NIC interrupts. If the interrupts are uneven, see CPU—frequency scaling, CPU—interrupt affinity, and NUMA CPU binding. If the NIC does not support RSS or the number of RSS is much smaller than the number of physical CPU cores, configure Receive Packet Steering (RPS, which can be regarded as the software implementation of RSS), and the RPS extension Receive Flow Steering (RFS). For detailed configuration, see the kernel document.

  • Software interrupts: Observe the monitoring of /proc/net/softnet_stat. If the values of the other columns except the third column are increasing, properly adjust the value of net.core.netdev_budget or net.core.dev_weight for softirq to get more CPU time. In addition, you also need to check the CPU usage to determine which tasks are frequently using the CPU and whether they can be optimized.

  • Receive queue of application sockets: Monitor the Resv-q column of ss -nmp. If the queue is full, consider increasing the size of the application socket cache or use the automatic cache adjustment method. In addition, consider whether you can optimize the architecture of the application layer and reduce the interval between reading sockets.

  • Ethernet flow control: If the NIC and switch support the flow control feature, you can use this feature to leave some time for the kernel to process the data in the NIC queue, to avoid the issue of NIC buffer overflow.

  • Interrupts coalescing: Too frequent hardware interrupts reduces system performance, and too late hardware interrupts causes packet loss. Newer NICs support the interrupt coalescing feature and allow the driver to automatically adjust the number of hardware interrupts. You can execute ethtool -c ${NIC_DEV_NAME} to check and ethtool -C ${NIC_DEV_NAME} to enable this feature. The adaptive mode allows the NIC to automatically adjust the interrupt coalescing. In this mode, the driver checks the traffic mode and kernel receiving mode, and evaluates the coalescing settings in real time to prevent packet loss. NICs of different brands have different features and default configurations. For details, see the NIC manuals.

  • Adapter queue: Before processing the protocol stack, the kernel uses this queue to buffer the data received by the NIC, and each CPU has its own backlog queue. The maximum number of packets that can be cached in this queue is netdev_max_backlog. Observe the second column of /proc/net/softnet_stat. When the second column of a row continues to increase, it means that the CPU [row-1] queue is full and the data packet is lost. To resolve this problem, continue to double the net.core.netdev_max_backlog value.

  • Send queue: The length value of a send queue determines the number of packets that can be queued before sending. The default value is 1000, which is sufficient for 10 Gbps. But if you have observed the value of TX errors from the output of ip -s link, you can try to double it: ip link set dev ${NIC_DEV_NAME} txqueuelen 2000.

  • Driver: NIC drivers usually provide tuning parameters. See the device hardware manual and its driver documentation.