:Differences between Alibaba Cloud Linux 3 and CentOS 7

sysctl system parameter

CentOS 7 x86

Alibaba Cloud Linux 3 x86 and Arm64

Description

Linux community patch information

Value

Read and write permissions

Value

Read and write permissions

kernel.panic

0

rw

1

rw

The default behaviors between CentOS 7 and Alibaba Cloud Linux 3 are different.

This option controls whether the kernel waits for n (the value of sysctl) seconds to automatically restart if a panic occurs. Valid values of n:

• n > 0: The kernel waits for n seconds and then restarts.

• n = 0: The kernel does not restart.

• n < 0: The kernel immediately restarts.

None

kernel.watchdog_thresh

10

rw

50

rw

The watchdog timeout periods (in seconds) between the two versions are different.

The cloud links are more complex. The value of the parameter is changed in Alibaba Cloud Linux to reduce the probability of false-positive warning alerts.

None

kernel.nmi_watchdog

1

rw

0

rw

The hard lockup detection states between the two versions are different.

Non-Maskable Interrupts (NMIs) are not supported in Elastic Compute Service (ECS), and therefore the hard lockup detection feature is not needed.

If you use bare metal devices and need the hard lockup detection feature, you can manually enable the hard lockup detection feature.

None

kernel.hung_task_timeout_secs

120

rw

240

rw

The timeout periods (in seconds), in which a process in the D (TASK_UNINTERRUPTIBLE) state is not scheduled, between the two versions are different.

The cloud links are more complex. The value of the parameter is changed in Alibaba Cloud Linux to reduce the probability of false-positive warning alerts.

None

vm.drop_caches

0

rw

None

-w

The read and write permissions between the two versions are different.

This option supports the behavior of reading the last written value in CentOS 7, which misleads users that this option is to specify the non-cached mode, instead of initiating a task to clear the cache. To prevent misunderstanding, the read permission is removed from Alibaba Cloud Linux 3.

• Commit ID: 204cb79ad42f015312a5bbd7012d09c93d9b46fb

• Subject information: kernel:sysctl: make drop_caches write-only

vm.dirty_ratio

30

rw

40

rw

The value of this option varies based on the profile or mode that is used to tune performance. In CentOS 7, the value of this option is set to 30 when the tuneD daemon uses the virtual-guest profile to tune performance. The value of this option is set to 40 in Alibaba Cloud Linux when the ECS-performance mode is used to tune performance.

None

vm.numa_zonelist_order

default

rw

Node

rw

The option becomes invalid during the iteration of the Linux community versions.

This option controls the order of selection from each node&zone during system memory allocation. Due to the iteration of the Linux community versions, manual control is no longer required.

• Commit ID: c9bff3eebc09be23fbc868f5e6731666d23cbea3

• Subject information: mm, page_alloc: rip out ZONELIST_ORDER_ZONE

net.ipv4.tcp_mem

None

rw

None

rw

This option consists of the following integer values:

• low: The system does not tune memory allocation if the number of pages allocated by TCP is less than the low value.

• pressure: The system reduces memory usage and enters the pressure mode if the number of pages allocated by TCP exceeds the pressure value.

• max: specifies the maximum number of pages that TCP can allocate to sockets.

The default value of this option varies based on the system memory size. However, for the same memory specification, the buffer size in Alibaba Cloud Linux 3 is about half of the buffer size in CentOS 7. This is because the initialization algorithm for default values was modified during iteration of the Linux community versions.

• Commit ID: b66e91ccbc34ebd5a2f90f9e1bc1597e2924a500

• Subject information: tcp: halvestcp_mem[] limit

net.ipv4.tcp_fastopen

0

rw

1

rw

The default value varies based on the Linux community version.

This option controls whether the system can send and receive messages related to TCP Fast Open. Valid values:

• 0: Messages related to TCP Fast Open cannot be sent or received.

• 1: Messages related to TCP Fast Open can only be sent.

• Commit ID: e1cfcbe82b4534bd0f99fef92a6d33843fd85e0e

• Subject information: ipv4: Namespaceify tcp_fastopen knob

net.ipv4.tcp_fack

1

rw

0

rw

The option becomes invalid during the iteration of the Linux community versions.

This option is used to support the TCP Forward Acknowledgment (FACK) feature. The FACK feature is replaced by the Recent Acknowledgment (RACK) feature. Consequently, this option becomes invalid in Alibaba Cloud Linux 3.

• Commit ID: 713bafea92920103cd3d361657406cf04d0e22dd

• Subject information: tcp: retire FACK loss detection

net.core.somaxconn

128

rw

4,096

rw

The default value varies based on the Linux community version.

This option is used to specify the maximum length of the socket listening queue. The maximum length of the socket listening queue is changed from 128 to 4,096 to meet the requirements of servers.

• Commit ID: 19f92a030ca6d772ab44b22ee6a01378a8cb32d4

• Subject information: net: increase SOMAXCONN to 4096

net.core.default_qdisc

pfifo_fast

rw

fq_codel

rw

The default Queuing Disciplines (qdiscs) models between the two versions are different. The qdiscs value is set to Fair Queuing Controlled Delay (FQ_Codel) in Alibaba Cloud Linux 3, which can prevent the bufferbloat issue.

None

sysctl system parameter

Value

Read and write permissions

Description

Linux community patch information

kernel.panic_on_stackoverflow

0

rw

The option was removed during the iteration of the Linux community versions.

This option is used to control kernel stack overflow detection. In subsequent kernel version iterations, the kernel stack has the guard pages feature to perform detection by default. Therefore, this option is invalid.

• Commit ID: 117ed45485413b1977bfc638c32bf5b01d53c62b

• Subject information: x86/irq/64: Remove stack overflow debug code

kernel.random.read_wakeup_threshold

64

rw

The option was removed during the iteration of the Linux community versions.

• Commit ID: c95ea0c69ffda19381c116db2be23c7e654dac98

• Subject information: random: remove kernel.random.read_wakeup_threshold

kernel.compat-log

1

rw

The option was removed during the iteration of the Linux community versions.

Virtual 8086 mode is not supported on x86 64-bit architectures. If this option is configured, warning logs are printed when vm86old() and vm86() are called. The unnecessary warning logs are removed. This option is also removed.

• Commit ID: 9dcfcda5768eda793e15a1a73da38cfd1fc1a47a

• Subject information: compat: remove compat_printk()

kernel.numa_balancing_settle_count

4

rw

The option was removed during the iteration of the Linux community versions.

Linux uses more scientific metrics to distinguish loads for different types of tasks. Therefore, this option is no longer used to control the frequency threshold.

• Commit ID: 1bd53a7efdc988163ec4c25f656df38dbe500632

• Subject information: sched/numa: Drop sysctl_numa_balancing_settle_count sysctl

kernel.sched_shares_window_ns

10,000,000

rw

The option was removed during the iteration of the Linux community versions. This feature is invalid in CentOS 7.

• Commit ID: 3c3fcb45d524feb5d14a14f332e3eec7f2aff8f3

• Subject information: sched/fair: Kill the unused 'sched_shares_window_ns' tunable

kernel.sched_time_avg_ms

1,000

rw

The option was removed during the iteration of the Linux community versions.

Linux uses the per-entity load tracking (PELT) algorithm for the real-time (RT) scheduling class. The predecessor algorithm and related parameters that include the option are no longer used.

• Commit ID: 5fd778915ad29184a5ff8eb82d1118f6916b79e4

• Subject information: sched/sysctl: Remove unused sched_time_avg_ms sysctl

vm.hugepages_treat_as_movable

0

rw

The option was removed during the iteration of the Linux community versions.

This option is intended to allow the allocation of non-migratable large pages from the ZONE_MOVABLE area to reduce memory fragments. This feature seriously affects the memory hot swapping feature and was therefore removed.

• Commit ID: d6cb41cc44c63492702281b1d329955ca767d399

• Subject information: mm, hugetlb: remove hugepages_treat_as_movable sysctl

vm.nr_pdflush_threads

0

r-

The option was removed during the iteration of the Linux community versions. This feature is invalid in CentOS 7.

• Commit ID: b35bd0d9f8a8ea17aae40893e18274d191a2d2c5

• Subject information: sysctl: remove /proc/sys/vm/nr_pdflush_threads

net.ipv4.tcp_tw_recycle

0

rw

The option was removed during the iteration of the Linux community versions.

This option allows the system to quickly reclaim sockets in the TIME-WAIT state. The feature is implemented based on the condition that the timestamps of the same remote address monotonically increase, whereas the timestamp random drift feature of NAT and TCP conflicts with the condition. As a result, the net.ipv4.tcp_tw_recycle option was removed because it cannot be used.

• Commit ID: 4396e46187ca5070219b81773c4e65088dac50cc

• Subject information: tcp: remove tcp_tw_recycle

net.ipv4.tcp_thin_dupack

0

rw

The option was removed during the iteration of the Linux community versions. This feature is replaced with the TCP RACK feature.

• Commit ID: 4a7f6009441144783e5925551c72e3f2e1b0839b

• Subject information: tcp: remove thin_dupack feature

net.ipv4.tcp_max_ssthresh

0

rw

The option was removed during the iteration of the Linux community versions. This option is intended to change the TCP slow start speed and is now replaced by a more effective and automated method. Therefore, the option is removed.

• Commit ID: 9f9843a751d0a2057f9f3d313886e7e5e6ebaac9

• Subject information: tcp: properly handle stretch acks in slow start

fs.negative-dentry-limit

0

rw

This option is an in-house feature developed in CentOS 7 and is not a common feature of the Linux community. The option is removed from CentOS 8.

Why negative-dentry-limit sysctl parameter is not available in Red Hat Enterprise Linux 8?

fs.may_detach_mounts

0

rw

This option is an in-house feature developed in CentOS 7 and is not a common feature of the Linux community. The option is removed from CentOS 8.

What is the parameter fs.may_detach_mounts on Red Hat Enterprise Linux 7?