cxlswap interface cannot be used properly
luohuang017 opened this issue · comments
When I tested it using test_cxlswap.c, I found that I was being oom killed. Upon inspection, it seems that there is an issue with the avoid_oom function in util.h . you are passing the -17 parameter to oom_score_adj instead of oom_adj. But even after I made the modification, the program still cannot use the cxlswap interface properly; instead, it just stops directly when reaching the memory limit.
Did you run test_cxlswap directly or using run_cxlswap_storeload_test.sh, etc?
If you run directly, then there is a possibility that CXL Swap Interface is not turned on.
If not turned on, then system just try to swap out to swap space (ssd, hdd, etc.).
("$ cat /sys/module/cxlswap/parameters/enabled" should true or y)
If CXL Swap Interface is already turned on, and if you run shared memory test, then can you try after decreasing the test size in run_cxlswap_sharedmemory_test.sh. It takes so long to finish that it appears as if it has stopped.
I also wonder if your system has swap space? (you can check it by "$ swapon -s")
As I know, what oom_adj for -17 means is just to request that it be given the lowest priority in the target of oom_kill.
But when the system is difficulty to operate PFRA cause it has no Swap Space, then your application will be killed frequently even the oom_adj_score is -17(below -16).
And, as I know the meaning of oom_adj and oom_score_adj is same, but oom_adj exists for backward compatilibty. (https://github.com/torvalds/linux/blob/master/Documentation/filesystems/proc.rst)
Did you run test_cxlswap directly or using run_cxlswap_storeload_test.sh, etc?
If you run directly, then there is a possibility that CXL Swap Interface is not turned on. If not turned on, then system just try to swap out to swap space (ssd, hdd, etc.). ("$ cat /sys/module/cxlswap/parameters/enabled" should true or y)
If CXL Swap Interface is already turned on, and if you run shared memory test, then can you try after decreasing the test size in run_cxlswap_sharedmemory_test.sh. It takes so long to finish that it appears as if it has stopped.
I also wonder if your system has swap space? (you can check it by "$ swapon -s")
As I know, what oom_adj for -17 means is just to request that it be given the lowest priority in the target of oom_kill. But when the system is difficulty to operate PFRA cause it has no Swap Space, then your application will be killed frequently even the oom_adj_score is -17(below -16). And, as I know the meaning of oom_adj and oom_score_adj is same, but oom_adj exists for backward compatilibty. (https://github.com/torvalds/linux/blob/master/Documentation/filesystems/proc.rst)
Thanks for your answer!
I found that I did not create the swap partition. So I created a swap file. But the program still doesn't work and I still can't figure out what's causing the problem.
Is the environment you run QEMU? In my experience, emulating and turning CXL in QEMU can operate very slowly and look like it has stopped.
And also, your system has CXL Memory in the form of Memory Only NUMA Node?
Is the environment you run QEMU? In my experience, emulating and turning CXL in QEMU can operate very slowly and look like it has stopped. And also, your system has CXL Memory in the form of Memory Only NUMA Node?
I run the program on my host.
Umm.. In my system, the test works well... and storeload test is very simple test...
- Did you run as root? Because this test needs cgroup interface (if system doesn't use cgroup before, then test try to mount cgroup interface and this operation needs root privilege. The version of cgroup isn't matter.)
- Can you check the number of "/sys/kernel/debug/cxlswap/stored_pages" is change while running test? (In my case, it chaged very frequently)
Umm.. In my system, the test works well... and storeload test is very simple test... Can you check the number of "/sys/kernel/debug/cxlswap/stored_pages" is change while running test? (In my case, it chaged very frequently)
I tried.But it didnot change:( It always be zero.
Oh, then can you check the number of reject_alloc_fail?
Can you check the THP option? CXLSwap is not allowed thp swap out case.
I think your case is that PFRA works and try to swap out by cxlswap, but allocation to cxl memory is fail and finally killed by OOM maybe...
Oh, then can you check the number of reject_alloc_fail? Can you check the THP option? CXLSwap is not allowed thp swap out case. I think your case is that PFRA works and try to swap out by cxlswap, but allocation to cxl memory is fail and finally killed by OOM maybe...
The file "/sys/kernel/mm/transparent_hugepage/enabled" was originally set to [madvise]. After I changed it to [always] and ran program again, it still failed. The "reject_alloc_fail" and "stored_pages" remain at 0 both before and after running it.The picture is the result about "dmesg | grep cxlswap".
Sorry for late response. Umm..
Maybe, the PFRA operations doesn't work at all...
I think there may be some errors in cgroup setting.. I will do some more tests.
Can you check the status of zswap of your system ("$ cat /sys/module/zswap/parameters/enabled")
And can you show me the status of /sys/module/cxlswap/parameters/*?
**SeungjunHa ** commented Mar 15, 2024 •
Thank you for response.
Can you turn off zswap($ echo N > /sys/module/zswap/parameters/enabled) and try again cxlswap_storeload_test?
Because zswap and cxlswap are frontswap api, if both are turned on simultaneously, zswap is used first in smdk kernel.
I think if you turned off zswap and turn on cxlswap, then the cxlswap's counters in debugfs are should be changed while doing test.
Thank you for response. Can you turn off zswap($ echo N > /sys/module/zswap/parameters/enabled) and try again cxlswap_storeload_test? Because zswap and cxlswap are frontswap api, if both are turned on simultaneously, zswap is used first in smdk kernel. I think if you turned off zswap and turn on cxlswap, then the cxlswap's counters in debugfs are should be changed while doing test.
It still be killed.
Umm.. can you check /proc/sys/vm/ variables like swappiness, overcommit_memory, panic_on_oom.
In my case,
$ cat /proc/sys/vm/swappiness
60 (If I set this value to zero, then I also got OOM Killed case.. I think the value of this variable is different..)
$ cat /proc/sys/vm/panic_on_oom
0
$ cat /proc/sys/vm/overcommit_memory
0
Umm.. can you check /proc/sys/vm/ variables like swappiness, overcommit_memory, panic_on_oom. In my case, $ cat /proc/sys/vm/swappiness 60 (If I set this value to zero, then I also got OOM Killed case.. I think the value of this variable is different..) $ cat /proc/sys/vm/panic_on_oom 0 $ cat /proc/sys/vm/overcommit_memory 0
That's my case, truely different from you.
Good. It may have been solved.
If you don't mind, I will close the thread.
Umm.. can you check /proc/sys/vm/ variables like swappiness, overcommit_memory, panic_on_oom. In my case, $ cat /proc/sys/vm/swappiness 60 (If I set this value to zero, then I also got OOM Killed case.. I think the value of this variable is different..) $ cat /proc/sys/vm/panic_on_oom 0 $ cat /proc/sys/vm/overcommit_memory 0
Good. It may have been solved. If you don't mind, I will close the thread.
Yes,thanks for your time.