On 2/26/20 2:26 PM, Johannes Weiner wrote:
On Wed, Feb 26, 2020 at 12:25:33PM -0800, Shakeel Butt wrote:
On Wed, Feb 19, 2020 at 10:12 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
We have received regression reports from users whose workloads moved
into containers and subsequently encountered new latencies. For some
users these were a nuisance, but for some it meant missing their SLA
response times. We tracked those delays down to cgroup limits, which
inject direct reclaim stalls into the workload where previously all
reclaim was handled my kswapd.
This patch adds asynchronous reclaim to the memory.high cgroup limit
while keeping direct reclaim as a fallback. In our testing, this
eliminated all direct reclaim from the affected workload.
memory.high has a grace buffer of about 4% between when it becomes
exceeded and when allocating threads get throttled. We can use the
same buffer for the async reclaimer to operate in. If the worker
cannot keep up and the grace buffer is exceeded, allocating threads
will fall back to direct reclaim before getting throttled.
For irq-context, there's already async memory.high enforcement. Re-use
that work item for all allocating contexts, but switch it to the
unbound workqueue so reclaim work doesn't compete with the workload.
The work item is per cgroup, which means the workqueue infrastructure
will create at maximum one worker thread per reclaiming cgroup.
Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
---
mm/memcontrol.c | 60 +++++++++++++++++++++++++++++++++++++------------
mm/vmscan.c | 10 +++++++--
This reminds me of the per-memcg kswapd proposal from LSFMM 2018
(https://lwn.net/Articles/753162/).
Ah yes, I remember those discussions. :)
One thing that has changed since we tried to implement this last was
the workqueue concurrency code. We don't have to worry about a single
thread or fixed threads per cgroup, because the workqueue code has
improved significantly to handle concurrency demands, and having one
work item per cgroup makes sure we have anywhere between 0 threads and
one thread per cgroup doing this reclaim work, completely on-demand.
Yes, exactly. Our in-house implementation was just converted to use
workqueue instead of dedicated kernel thread for each cgroup.
Also, with cgroup2, memory and cpu always have overlapping control
domains, so the question who to account the work to becomes a much
easier one to answer.
If I understand this correctly, the use-case is that the job instead
of direct reclaiming (potentially in latency sensitive tasks), prefers
a background non-latency sensitive task to do the reclaim. I am
wondering if we can use the memory.high notification along with a new
memcg interface (like memory.try_to_free_pages) to implement a user
space background reclaimer. That would resolve the cpu accounting
concerns as the user space background reclaimer can share the cpu cost
with the task.
The idea is not necessarily that the background reclaimer is lower
priority work, but that it can execute in parallel on a separate CPU
instead of being forced into the execution stream of the main work.
So we should be able to fully resolve this problem inside the kernel,
without going through userspace, by accounting CPU cycles used by the
background reclaim worker to the cgroup that is being reclaimed.
Actually I'm wondering if we really need account CPU cycles used by
background reclaimer or not. For our usecase (this may be not general),
the purpose of background reclaimer is to avoid latency sensitive
workloads get into direct relcaim (avoid the stall from direct relcaim).
In fact it just "steal" CPU cycles from lower priority or best-effort
workloads to guarantee latency sensitive workloads behave well. If the
"stolen" CPU cycles are accounted, it means the latency sensitive
workloads would get throttled from somewhere else later, i.e. by CPU share.
We definitely don't want to the background reclaimer eat all CPU cycles.
So, the whole background reclaimer is opt in stuff. The higher level
cluster management and administration components make sure the cgroups
are setup correctly, i.e. enable for specific cgroups, setup watermark
properly, etc.
Of course, this may be not universal and may be just fine for some
specific configurations or usecases.
One concern with this approach will be that the memory.high
notification is too late and the latency sensitive task has faced the
stall. We can either introduce a threshold notification or another
notification only limit like memory.near_high which can be set based
on the job's rate of allocations and when the usage hits this limit
just notify the user space.
Yeah, I think it would be a pretty drastic expansion of the memory
controller's interface.
