When the Linux operating system is up for an extended period of time, memory fragmentation may occur. This topic provides solutions to this issue.

Problem description

When a service deployed on an instance occasionally takes an extended period of time responding or when a system call takes an extended period of time, the sys metric value of the system increases accordingly. The buddy system is in lack of high-order memory (memory whose order is higher than 3). The following code shows an example output of the cat /proc/buddyinfo command. Each column starting from the fourth column indicates the free memory at a different order of the buddy system. 
cat /proc/buddyinfo 
Node 0, zone      DMA      1      0      0   1   2    1    1     0     1      1     3 
Node 0, zone    DMA32   3173    856    529   0   0    0    0     0     0      0     0 
Node 0, zone   Normal  19030   8688   7823   0   0    0    0     0     0      0     0

Possible causes

When the Linux operating system is up for an extended period of time, contiguous large physical memory chunks are split into small physical memory blocks. If a service requires contiguous large memory chunks, the system starts a memory compaction procedure that takes time, which may cause system performance jitters. Memory fragmentation typically generates kernel stack information. The following code shows an example of the kernel stack information: 
0xffffffff8118f9cb compaction_alloc  ([kernel.kallsyms])
0xffffffff811c88a9 migrate_pages  ([kernel.kallsyms])
0xffffffff811901ee compact_zone  ([kernel.kallsyms])
0xffffffff8119041b compact_zone_order  ([kernel.kallsyms])
0xffffffff81190735 try_to_compact_pages  ([kernel.kallsyms])
0xffffffff81631cb4 __alloc_pages_direct_compact  ([kernel.kallsyms])
0xffffffff811741d5 __alloc_pages_nodemask  ([kernel.kallsyms])
0xffffffff811b5a79 alloc_pages_current  ([kernel.kallsyms])
0xffffffff811c0005 new_slab  ([kernel.kallsyms])
0xffffffff81633848 __slab_alloc  ([kernel.kallsyms])
0xffffffff811c5291 __kmalloc_node_track_caller  ([kernel.kallsyms])
0xffffffff8151a8c1 __kmalloc_reserve.isra.30  ([kernel.kallsyms])
0xffffffff8151b7cd alloc_sib  ([kernel.kallsyms])
0xffffffff815779e9 sk_stream_alloc_skb  ([kernel.kallsyms])
0xffffffff8157872d tcp_sendmsg  ([kernel.kallsyms])
0xffffffff815a26b4 inet_sendmsg  ([kernel.kallsyms])
0xffffffff81511017 sock_aio_write  ([kernel.kallsyms])
0xffffffff811df729 do_sync_readv_writev  ([kernel.kallsyms])
0xffffffff811e0cfe do_readv_writev  ([kernel.kallsyms])

Solutions

You can take one of the following measures to troubleshoot memory fragmentation in the Linux operating system:
  • Adjust the min watermark
    In most cases, we recommend that you set the min watermark to 1% to 3% of the total memory. We recommend that you set it to 2% of the total memory. When memory resources are insufficient, asynchronous reclaim is triggered. You can run the following command to adjust the min watermark:
    sysctl -w vm.min_free_kbytes = memtotal_kbytes * 2%
    In the preceding command, the memtotal_kbytes * 2% variable indicates the memory size that is 2% of the total memory of the instance.
  • Adjust the difference between the min and low watermarks
    You can specify the watermark_scale_factor kernel parameter to adjust the difference between the min and low watermarks to cope with sudden memory demands. The default value of watermark_scale_factor is 0.1% of the total memory, and the minimum value is calculated based on the following formula: 0.5 × <min watermark>. The minimum value is the minimum difference between the min and low watermarks. You can run the following command to adjust the value of watermark_scale_factor:
    sysctl -w vm.watermark_scale_factor = value
    In the preceding command, the value variable indicates your specified difference between the min and low watermarks.
  • Perform regular memory compaction
    You can run the following command to trigger asynchronous memory compaction during off-peak hours: 
    echo 1 > /proc/sys/vm/compact_memory
  • Manually drop the cache at regular intervals
    If the preceding measures cannot effectively handle memory fragmentation, you can also drop the cache during off-peak hours so that the memory can be re-allocated. You can effectively prevent memory fragmentation by dropping the cache. However, short-period system performance jitters may occur when you perform this operation. You can run the following command to drop the cache:
    echo 3 > /proc/sys/vm/drop_caches