chmem

Set a memory block offline

$ sudo chmem [[-b|--block]] [[-d|--disable]] [block_number]

$ sudo chmem [[-b|--block]] [[-e|--enable]] [block_number]

$ sudo chmem [[-d|--disable]] 0x[start_address]-0x[end_address]

$ sudo chmem [[-e|--enable]] 0x[start_address]-0x[end_address]

$ sudo chmem [[-e|--enable]] 0x[start_address] [[-z|--zone]] [Movable]

$ chmem [[-h|--help]]

chmem [options] <pid>

--hugepages
    Set fixed number of huge pages for the process

--hugepages-min
    Set minimum number of huge pages allowed

--hugepages-max
    Set maximum number of huge pages allowed

--hugepagesize
    Specify huge page size (2M or 1G, default from kernel)

--mpol-default
    Set default memory policy

--mpol-prefer
    Prefer memory allocation from specific NUMA node

--mpol-bind
    Bind allocations to comma-separated list of NUMA nodes

--mpol-interleave
    Interleave allocations across NUMA nodes

--mpol-local
    Prefer local NUMA node for allocations

--ismemorypolicy
    Enable (yes) or disable (no) memory policy enforcement (default: yes)

--help
    Display help message

--version
    Show version information

chmem is a utility from the util-linux package that modifies the memory policy or huge page allocation settings for a running process. It allows administrators to adjust how a process allocates memory, particularly in NUMA (Non-Uniform Memory Access) systems or when using huge pages for performance optimization.

Huge pages reduce TLB (Translation Lookaside Buffer) overhead by using larger memory pages (e.g., 2MB or 1GB instead of 4KB), improving performance for memory-intensive applications like databases or virtual machines. chmem can set the number of huge pages a process is allowed to use, either as a fixed amount, minimum, or maximum.

Additionally, it supports setting NUMA memory policies, such as preferring local memory, binding to specific nodes, or interleaving across nodes. This is crucial on multi-socket systems where memory access latency varies by node. Policies can be default, prefer, bind, interleave, or local.

The command requires root privileges and operates on a specified PID. Changes apply immediately but may not retroactively affect already allocated memory. It's useful for tuning workloads dynamically without restarting processes. Note that support depends on kernel features like CONFIG_TRANSPARENT_HUGEPAGE and NUMA awareness.

Requires root privileges; changes may not affect existing allocations; kernel must support huge pages and NUMA policies; invalid options or PID cause failure without warning.

chmem --hugepages=1024 1234
Sets 1024 huge pages for PID 1234.

chmem --mpol-prefer=1 --ismemorypolicy=yes 5678
Prefer NUMA node 1 for PID 5678.

Introduced in util-linux 2.27 (2014); enhanced in later versions for better NUMA support and huge page options.

numactl(1), taskset(1), mbind(2)