Re: [RFC 0/7] Initial proposal for faster res_counter updates [message #45717 is a reply to message #45710] |
Fri, 30 March 2012 08:32 ![Go to previous message Go to previous message](/theme/ovz3/images/up.png) ![Go to next message Go to previous message](/theme/ovz3/images/down.png) |
KAMEZAWA Hiroyuki
Messages: 463 Registered: September 2006
|
Senior Member |
|
|
(2012/03/30 17:04), Glauber Costa wrote:
> Hi,
>
> Here is my take about how we can make res_counter updates faster.
> Keep in mind this is a bit of a hack intended as a proof of concept.
>
> The pros I see with this:
>
> * free updates in non-constrained paths. non-constrained paths includes
> unlimited scenarios, but also ones in which we are far from the limit.
>
> * No need to have a special cache mechanism in memcg. The problem with
> the caching is my opinion, is that we will forward-account pages, meaning
> that we'll consider accounted pages we never used. I am not sure
> anyone actually ran into this, but in theory, this can fire events
> much earlier than it should.
>
Note: Assume a big system which has many cpus, and user wants to devide
the system into containers. Current memcg's percpu caching is done
only when a task in memcg is on the cpu, running. So, it's not so dangerous
as it looks.
But yes, if we can drop memcg's code, it's good. Then, we can remove some
amount of codes.
> But the cons:
>
> * percpu counters have signed quantities, so this would limit us 4G.
> We can add a shift and then count pages instead of bytes, but we
> are still in the 16T area here. Maybe we really need more than that.
>
....
struct percpu_counter {
raw_spinlock_t lock;
s64 count;
s64 limtes us 4G ?
> * some of the additions here may slow down the percpu_counters for
> users that don't care about our usage. Things about min/max tracking
> enter in this category.
>
I think it's not very good to increase size of percpu counter. It's already
very big...Hm. How about
struct percpu_counter_lazy {
struct percpu_counter pcp;
extra information
s64 margin;
}
?
> * growth of the percpu memory.
>
This may be a concern.
I'll look into patches.
Thanks,
-Kame
> It is still not clear for me if we should use percpu_counters as this
> patch implies, or if we should just replicate its functionality.
>
> I need to go through at least one more full round of auditing before
> making sure the locking is safe, specially my use of synchronize_rcu().
>
> As for measurements, the cache we have in memcg kind of distort things.
> I need to either disable it, or find the cases in which it is likely
> to lose and benchmark them, such as deep hierarchy concurrent updates
> with common parents.
>
> I also included a possible optimization that can be done when we
> are close to the limit to avoid the initial tests altogether, but
> it needs to be extended to avoid scanning the percpu areas as well.
>
> In summary, if this is to be carried forward, it definitely needs
> some love. It should be, however, more than enough to make the
> proposal clear.
>
> Comments are appreciated.
>
> Glauber Costa (7):
> split percpu_counter_sum
> consolidate all res_counter manipulation
> bundle a percpu counter into res_counters and use its lock
> move res_counter_set limit to res_counter.c
> use percpu_counters for res_counter usage
> Add min and max statistics to percpu_counter
> Global optimization
>
> include/linux/percpu_counter.h | 3 +
> include/linux/res_counter.h | 63 ++++++-----------
> kernel/res_counter.c | 151 +++++++++++++++++++++++++++++-----------
> lib/percpu_counter.c | 16 ++++-
> 4 files changed, 151 insertions(+), 82 deletions(-)
>
|
|
|