Basic Component Parameters

kernel.numa_balancing

Automatic NUMA balancing

{0, 1}

-

If CPU core binding is not required, set this parameter to 1.

Viewing: sysctl param

Setting: sysctl param=value

kernel.sched_child_runs_first

CPU scheduling priority of a subprocess

{0, 1}

TensorFlow Serving

-

kernel.sched_cluster

Cluster scheduling affinity

{0, 1}

MySQL, Redis, TensorFlow Serving

Set this parameter to 1 to enable cluster scheduling affinity.

kernel.sched_migration_cost_ns

Minimum cost for process migration

[100000, 5000000]

TensorFlow Serving

-

kernel.sched_nr_migrate

Number of tasks migrated by the scheduler during load balancing

[1, 128]

TensorFlow Serving

-

kernel.sched_wakeup_granularity_ns

Scheduling delay threshold for woken tasks

[100000, 15000000]

TensorFlow Serving

-

max_sectors_kb

Data size of a single I/O operation on a block device

[68, 128]

MySQL

For sequential read/write-intensive services, set this parameter to a larger value to increase the throughput. For random I/O latency-sensitive services, set this parameter to a smaller value to reduce the response time.

Viewing or modifying the file content:

/sys/block/block_device/queue/max_sectors_kb

scheduler

Block device I/O scheduling policy

{mq-deadline, kyber,bfq, none}

MySQL, openGauss

Set this parameter based on the application type.

Viewing or modifying the file content:

/sys/block/block_device/queue/scheduler

transparent_hugepage_mode

Transparent huge page status

{madvise, never, always}

Redis, Hive, Spark, Kafka, TensorFlow Serving

Set this parameter to always.

Viewing or modifying the file content:

/sys/kernel/mm/transparent_hugepage/enabled

vm.dirty_background_ratio

Maximum percentage of memory that can become dirty before background flushing of dirty pages to drives starts, relative to the total memory size

[0, 100]

RocksDB

In latency-sensitive application scenarios, increase the parameter value to reduce the drive flushing frequency and latency fluctuation.

Viewing: sysctl param

Setting: sysctl param=value

vm.max_map_count

Maximum number of memory mapping (mmap) regions that can be created by a single process

[65530, 10000000]

RocksDB

Increase the number of mmap regions as required.

vm.vfs_cache_pressure

Tendency of reclaiming the memory which is used for caching of directory and inode objects (VFS caches)

[0, 500]

RocksDB

To prevent the caches from occupying too much memory, increase the parameter value when the memory space becomes insufficient.

net.core.rps_sock_flow_entries

Size of the Receive Packet Steering (RPS) hash table

[0, 131072]

-

Set this parameter to a relatively large value.

Viewing: sysctl param

Setting: sysctl param=value

AlwaysPreTouch

Indicates whether the JVM accesses and submits the heap memory to the physical memory during JVM initialization

{on, off}

• on: -XX:+param
• off: -XX:-param

G1HeapRegionSize

Minimum region size of the G1 heap

{1m, 2m, 4m, 8m, 16m, 32m}

-XX:param=value

G1MaxNewSizePercent

Maximum percentage of the G1 young generation in the heap

[20, 40]

-XX:param=value

Set the parameter together with -XX:+UnlockExperimentalVMOptions.

G1NewSizePercent

Initial percentage of the G1 young generation in the heap

[20, 40]

G1UseAdaptiveIHOP

Indicates whether to enable adaptive Initiating Heap Occupancy Percent (IHOP)

{on, off}

• on: -XX:+param
• off: -XX:-param

Set the parameter together with -XX:+UnlockExperimentalVMOptions.

MaxGCPauseMillis

Target for the maximum GC pause time

[200, 500]

-XX:param=value

MaxJavaStackTraceDepth

Maximum depth of an abnormal stack trace

[512, 1024]

ParallelGCThreads

Number of STW worker threads

[16, 128]

UseNUMA

Indicates whether to enable NUMA optimization

{on, off}

• on: -XX:+param
• off: -XX:-param