1

We are currently testing tiered memory demotion on a machine equipped with a CXL device.

To facilitate this, we created a specific script (https://github.com/hyun-sa/comem) and are using the memory.high setting within a cgroup to force memory demotion.

These are the commands we used to enable demotion:

echo 1 > /sys/kernel/mm/numa/demotion_enabled echo 2 > /proc/sys/kernel/numa_balancing 

The issue we're facing is that while demotion does occur, it proceeds extremely slowly—even slower than swapping to disk. Furthermore, during a 7-Zip benchmark, we observe a severe drop in CPU utilization, as if some process is causing a lock.

This is our running example (7zr b -md25 while memory is limited as 8G by memory.high)

7-Zip (r) 23.01 (x64) : Igor Pavlov : Public domain : 2023-06-20 64-bit locale=C.UTF-8 Threads:128 OPEN_MAX:1024 d25 Compiler: 13.2.0 GCC 13.2.0: SSE2 Linux : 6.15.6 : #1 SMP PREEMPT_DYNAMIC Tue Jul 15 06:39:48 UTC 2025 : x86_64 PageSize:4KB THP:madvise hwcap:2 hwcap2:2 AMD EPYC 9554 64-Core Processor (A10F11) 1T CPU Freq (MHz): 3710 3731 3732 3733 3733 3732 3732 64T CPU Freq (MHz): 6329% 3674 6006% 3495 RAM size: 386638 MB, # CPU hardware threads: 128 RAM usage: 28478 MB, # Benchmark threads: 128 Compressing | Decompressing Dict Speed Usage R/U Rating | Speed Usage R/U Rating KiB/s % MIPS MIPS | KiB/s % MIPS MIPS 22: 477942 10925 4256 464943 | 5843081 12451 4001 498193 23: 337115 8816 3896 343480 | 5826376 12606 3999 504053 24: 1785 108 1772 1919 | 5654618 12631 3928 496161 25: 960 63 1739 1097 | 1767869 4606 3415 157287 ---------------------------------- | ------------------------------ Avr: 204451 4978 2916 202860 | 4772986 10573 3836 413924 Tot: 7776 3376 308392 execution_time(ms): 2807639 

Is there a potential misunderstanding of how cgroups function or a misconfiguration in my setup that could be causing this behavior?

Our machine specifications are as follows:

Mainboard : Supermicro H13SSL-NT

CPU : Epyc 9554 (nps 1)

Dram : 128G

CXL device : SMART Modular Technologies Device c241

OS : Ubuntu 24.04 LTS

Kernel : Linux 6.15.6

numactl -H available: 2 nodes (0-1) node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 node 0 size: 128640 MB node 0 free: 117909 MB node 1 cpus: node 1 size: 257998 MB node 1 free: 257840 MB node distances: node 0 1 0: 10 50 1: 255 10 

Thank you for your help.

1
  • It's surprising to me you don't at least have comments. But note the low number of followers for all of your tags. If you could cobble together a mcve that would make it a programming problem suitable for StackOverflow. Add the mcve and repost there. ... AND/OR look at what other tags on U&L you could swap to that have much larger follower counts. (I have not relevant experience in your problem area) Commented Jul 25 at 23:18

0

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.