memcg: Add swappiness to cgroup2
Subsystems (3)
RFC 1/1 memcg: Add swappiness to cgroup2
Even if cgroup2 has swap.max, swappiness is still a very useful config. This commit add swappiness to cgroup2.
--- mm/memcontrol.c | 5 +++++ 1 file changed, 5 insertions(+)
mm/memcontrol.c +5 -0
@@ -7143,6 +7143,11 @@ static struct cftype swap_files[] = { .file_offset = offsetof(struct mem_cgroup, swap_events_file), .seq_show = swap_events_show, }, + { + .name = "swappiness", + .read_u64 = mem_cgroup_swappiness_read, + .write_u64 = mem_cgroup_swappiness_write, + }, { }/* terminate */ };
Signature · 1 line
2.7.4
Hi Hui,
Quoted 2 lines · Hui Zhu
Even if cgroup2 has swap.max, swappiness is still a very useful config. This commit add swappiness to cgroup2.
When submitting patches like this, it's important to explain *why* you want it and what evidence there is. For example, how should one use this to compose a reasonable system? Why aren't existing protection controls sufficient for your use case? Where's the data?
Also, why would swappiness be something cgroup-specific instead of hardware-specific, when desired swappiness is really largely about the hardware you have in your system?
I struggle to think of situations where per-cgroup swappiness would be useful, since it's really not a workload-specific setting.
Thanks,
Chris
Quoted 25 lines
Signed-off-by Hui Zhu <[email protected]>---
mm/memcontrol.c| 5 +++++1 file changed, 5 insertions(+)mm/memcontrol.c +5
-0@@ -7143,6 +7143,11 @@ static struct cftype swap_files[] = { .file_offset = offsetof(struct mem_cgroup, swap_events_file), .seq_show = swap_events_show, }, + { + .name = "swappiness", + .read_u64 = mem_cgroup_swappiness_read, + .write_u64 = mem_cgroup_swappiness_write, + }, { }/* terminate */ };Signature · 1 line
2.7.4
Quoted 7 lines · 在 2019年12月25日,22:05,Chris Down <[email protected]> 写道:
Hi Hui,
Hui Zhu writes:
Quoted 8 lines
Quoted 2 lines
Even if cgroup2 has swap.max, swappiness is still a very useful config. This commit add swappiness to cgroup2.
When submitting patches like this, it's important to explain *why* you want it and what evidence there is. For example, how should one use this to compose a reasonable system? Why aren't existing protection controls sufficient for your use case? Where's the data?
Also, why would swappiness be something cgroup-specific instead of hardware-specific, when desired swappiness is really largely about the hardware you have in your system?
I struggle to think of situations where per-cgroup swappiness would be useful, since it's really not a workload-specific setting.
Hi Chris,
My thought about per-cgroup swappiness is different applications should have different memory footprint.
For example, an application does a lot of file access work in a memory-constrained environment. Its performance depend on the file access speed. Keep more file cache will good for it. Then more swappiness will good for it, especially with the high speed swap device(zram/zswap). And in the same environment, an application that access anon memory a lot of times. Use low swapiness will good for its performance. But just let it not to swap is not a good for it because the code is inside file cache. Just drop the file cache will decrease the application speed sometime. Both of them are extreme examples. Other applications maybe access both file and anon. Maybe define a special swapiness is good for it.
This is what I thought about add swappiness to cgroup2.
Best,
Hui
Quoted 29 lines
Thanks,
Chris
Quoted 24 lines
Signed-off-by Hui Zhu <[email protected]>---
mm/memcontrol.c| 5 +++++1 file changed, 5 insertions(+)mm/memcontrol.c +5
-0@@ -7143,6 +7143,11 @@ static struct cftype swap_files[] = { .file_offset = offsetof(struct mem_cgroup, swap_events_file), .seq_show = swap_events_show, }, + { + .name = "swappiness", + .read_u64 = mem_cgroup_swappiness_read, + .write_u64 = mem_cgroup_swappiness_write, + }, { }/* terminate */ };Signature · 1 line
2.7.4
Michal Koutný
maintainerQuoted 4 lines · teawater
For example, an application does a lot of file access work in a memory-constrained environment.
[...]
Both of them are extreme examples.
The examples are quite generic. Do cgroup v2 controls really prevent handling such workloads appropriately?
Besides that, note that per-cgroup swappiness as used in v1 cannot be simply transferred into v2 because, it's a concept that doesn't take into account cgroup hierarchies (how would parent's swappiness affect children? what would swappiness on inner nodes mean?).
HTH,
Michal
Michal Hocko
maintainerQuoted 10 lines · Chris Down
Hi Hui,
Quoted 2 lines · Hui Zhu
Even if cgroup2 has swap.max, swappiness is still a very useful config. This commit add swappiness to cgroup2.
When submitting patches like this, it's important to explain *why* you want it and what evidence there is. For example, how should one use this to compose a reasonable system? Why aren't existing protection controls sufficient for your use case? Where's the data?
Agreed!
Quoted 3 lines
Also, why would swappiness be something cgroup-specific instead of hardware-specific, when desired swappiness is really largely about the hardware you have in your system?
I am not really sure I agree here though. Swappiness has been traditionally more about workload because it has been believed that it is a preference of the workload whether the anonymous or disk based memory is more important. Whether this is a good interface is debatable of course but time has shown that it is extremely hard to tune.
Not to mention that swappiness has been ignored for years for vast majority workloads because of the highly biased file LRU reclaim.
At the time when cgroup v2 was introduced it'd been claimed that we do not want to copy the v1 swappiness logic because of the semantic shortcomings and that a better tuning should developed in future replacing even the global knob. AFAIR Johannes wanted to have a refault vs. cost based file/anon balancing.
The lack of a sensible hierarchical behavior has been even a stronger argument.
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 5 lines · Michal Hocko
I am not really sure I agree here though. Swappiness has been traditionally more about workload because it has been believed that it is a preference of the workload whether the anonymous or disk based memory is more important. Whether this is a good interface is debatable of course but time has shown that it is extremely hard to tune.
Sure, it can theoretically be hardware- and workload-specific -- I don't think we disagree here. The reason I suggest it's a generally hardware-specific tunable rather than a workload-specific tunable is it's pretty rare to see anyone who's meaningfully used it for workload-specific tuning :-)
Andrew Morton
maintainerQuoted 8 lines · On Wed, 25 Dec 2019 13:36:48 +0800 Hui Zhu <[email protected]> wrote:
Even if cgroup2 has swap.max, swappiness is still a very useful config. This commit add swappiness to cgroup2.
...
mm/memcontrol.c static struct cftype swap_files[] = { · line 7152
Quoted 10 lines
mm/memcontrol.c +5
-0@@ -7143,6 +7143,11 @@ static struct cftype swap_files[] = { .file_offset = offsetof(struct mem_cgroup, swap_events_file), .seq_show = swap_events_show, }, + { + .name = "swappiness", + .read_u64 = mem_cgroup_swappiness_read, + .write_u64 = mem_cgroup_swappiness_write, + }, { }/* terminate */ };
There should be a Documentation/ update with this?
Quoted 11 lines · 在 2020年1月1日,07:16,Andrew Morton <[email protected]> 写道:
Quoted 5 lines · On Wed, 25 Dec 2019 13:36:48 +0800 Hui Zhu <[email protected]> wrote:
Even if cgroup2 has swap.max, swappiness is still a very useful config. This commit add swappiness to cgroup2.
...
Quoted 16 lines
mm/memcontrol.c static struct cftype swap_files[] = { · line 7152
Quoted 13 lines
mm/memcontrol.c +5
-0@@ -7143,6 +7143,11 @@ static struct cftype swap_files[] = { .file_offset = offsetof(struct mem_cgroup, swap_events_file), .seq_show = swap_events_show, }, + { + .name = "swappiness", + .read_u64 = mem_cgroup_swappiness_read, + .write_u64 = mem_cgroup_swappiness_write, + }, { }/* terminate */ };There should be a
Documentation/update with this?
Hi Andrew,
I will post a new version with Documentation/ update.
Thanks,
Hui