MOT Server Optimization – x86

Generally, databases are bounded by the following components –

• CPU – A faster CPU speeds up any CPU-bound database.
• Disk – High-speed SSD/NVME speeds up any I/O-bound database.
• Network – A faster network speeds up any SQL\Net-bound database.

In addition to the above, the following general-purpose server settings are used by default and may significantly affect a database's performance.

MOT performance tuning is a crucial step for ensuring fast application functionality and data retrieval. MOT can utilize state-of-the-art hardware, and therefore it is extremely important to tune each system in order to achieve maximum throughput.

The following are optional settings for optimizing MOT database performance running on an Intel x86 server. These settings are optimal for high throughput workloads –

BIOS

• Hyper Threading – ON

  Activation (HT=ON) is highly recommended.

  We recommend turning hyper threading ON while running OLTP workloads on MOT. When hyper‑threading is used, some OLTP workloads demonstrate performance gains of up to40%.

OS Environment Settings

• NUMA

  Disable NUMA balancing, as described below. MOT performs its own memory management with extremely efficient NUMA-awareness, much more than the default methods used by the operating system.

  echo 0 > /proc/sys/kernel/numa_balancing
  

• Services

  Disable Services, as described below –

  service irqbalance stop           # MANADATORY
  service sysmonitor stop           # OPTIONAL, performance 
  service rsyslog stop       # OPTIONAL, performance
  

• Tuned Service

  The following section is mandatory.

  The server must run the throughput-performance profile –

  [...]$ tuned-adm profile throughput-performance 
  

  The throughput-performance profile is broadly applicable tuning that provides excellent performance across a variety of common server workloads.

  Other less suitable profiles for openGauss and MOT server that may affect MOT's overall performance are – balanced, desktop, latency-performance, network-latency, network-throughput and powersave.

• Sysctl

  The following lists the recommended operating system settings for best performance.

  • Add the following settings to /etc/sysctl.conf and run sysctl -p

    net.ipv4.ip_local_port_range = 9000 65535
    kernel.sysrq = 1
    kernel.panic_on_oops = 1
    kernel.panic = 5
    kernel.hung_task_timeout_secs = 3600
    kernel.hung_task_panic = 1
    vm.oom_dump_tasks = 1
    kernel.softlockup_panic = 1
    fs.file-max = 640000
    kernel.msgmnb = 7000000
    kernel.sched_min_granularity_ns = 10000000
    kernel.sched_wakeup_granularity_ns = 15000000
    kernel.numa_balancing=0
    vm.max_map_count = 1048576
    net.ipv4.tcp_max_tw_buckets = 10000
    net.ipv4.tcp_tw_reuse = 1
    net.ipv4.tcp_tw_recycle = 1
    net.ipv4.tcp_keepalive_time = 30
    net.ipv4.tcp_keepalive_probes = 9
    net.ipv4.tcp_keepalive_intvl = 30
    net.ipv4.tcp_retries2 = 80
    kernel.sem = 250 6400000 1000 25600
    net.core.wmem_max = 21299200
    net.core.rmem_max = 21299200
    net.core.wmem_default = 21299200
    net.core.rmem_default = 21299200
    #net.sctp.sctp_mem = 94500000 915000000 927000000
    #net.sctp.sctp_rmem = 8192 250000 16777216
    #net.sctp.sctp_wmem = 8192 250000 16777216
    net.ipv4.tcp_rmem = 8192 250000 16777216
    net.ipv4.tcp_wmem = 8192 250000 16777216
    net.core.somaxconn = 65535
    vm.min_free_kbytes = 26351629
    net.core.netdev_max_backlog = 65535
    net.ipv4.tcp_max_syn_backlog = 65535
    #net.sctp.addip_enable = 0
    net.ipv4.tcp_syncookies = 1
    vm.overcommit_memory = 0
    net.ipv4.tcp_retries1 = 5
    net.ipv4.tcp_syn_retries = 5
    

  • Update the section of /etc/security/limits.conf to the following –

    <user> soft nofile 100000
    <user> hard nofile 100000
    

  The soft and a hard limit settings specify the quantity of files that a process may have opened at once. The soft limit may be changed by each process running these limits up to the hard limit value.

• Disk/SSD

  The following describes how to ensure that disk R/W performance is suitable for database synchronous commit mode.

  To do so, test your disk bandwidth using the following

  [...]$ sync; dd if=/dev/zero of=testfile bs=1M count=1024; sync
  1024+0 records in
  1024+0 records out
  1073741824 bytes (1.1 GB) copied, 1.36034 s, 789 MB/s 
  

  In case the disk bandwidth is significantly below the above number (789 MB/s), it may create a performance bottleneck for openGauss, and especially for MOT.

Network

Use a 10Gbps network or higher.

To verify, use iperf, as follows –

Server side: iperf -s
Client side: iperf -c <IP>

• rc.local – Network Card Tuning

  The following optional settings have a significant effect on performance –

  1. Copy set_irq_affinity.sh from https://gist.github.com/SaveTheRbtz/8875474 to /var/scripts/.

  2. Put in /etc/rc.d/rc.local and run chmod in order to ensure that the following script is executed during boot –

     'chmod +x /etc/rc.d/rc.local' 
     var/scripts/set_irq_affinity.sh -x all <DEVNAME>
     ethtool -K <DEVNAME> gro off
     ethtool -C <DEVNAME> adaptive-rx on adaptive-tx on
     Replace <DEVNAME> with the network card, i.e. ens5f1